Dom/OmniParser
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #116 from ThomasDh-C/demo

Browse Source 
Init gradio demo of computer use
This commit is contained in:
yadong-lu
2025-02-06 17:59:22 -06:00
committed by GitHub
parent ebc3912727 5ba2a7f144
commit 5d3a97db4f
73 changed files with 5328 additions and 4059 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
.gitignore vendored
View File

@@ -6,3 +6,7 @@ weights/icon_detect_v1_5_2/
.gradio
__pycache__/
debug.ipynb
util/__pycache__/
index.html?linkid=2289031
wget-log
weights/omniv2/

201
Dockerfile
View File

@@ -1,201 +0,0 @@
# Dockerfile for OmniParser with GPU support and OpenGL libraries
#
# This Dockerfile is intended to create an environment with NVIDIA CUDA
# support and the necessary dependencies to run the OmniParser project.
# The configuration is designed to support applications that rely on
# Python 3.12, OpenCV, Hugging Face transformers, and Gradio. Additionally,
# it includes steps to pull large files from Git LFS and a script to
# convert model weights from .safetensor to .pt format. The container
# runs a Gradio server by default, exposed on port 7861.
#
# Base image: nvidia/cuda:12.3.1-devel-ubuntu22.04
#
# Key features:
# - System dependencies for OpenGL to support graphical libraries.
# - Miniconda for Python 3.12, allowing for environment management.
# - Git Large File Storage (LFS) setup for handling large model files.
# - Requirement file installation, including specific versions of
# OpenCV and Hugging Face Hub.
# - Entrypoint script execution with Gradio server configuration for
# external access.
FROM nvidia/cuda:12.3.1-devel-ubuntu22.04
# Install system dependencies with explicit OpenGL libraries
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
git \
git-lfs \
wget \
libgl1 \
libglib2.0-0 \
libsm6 \
libxext6 \
libxrender1 \
libglu1-mesa \
libglib2.0-0 \
libsm6 \
libxrender1 \
libxext6 \
python3-opencv \
&& apt-get clean \
&& rm -rf /var/lib/apt/lists/* \
&& git lfs install
# Install Miniconda for Python 3.12
RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh && \
bash miniconda.sh -b -p /opt/conda && \
rm miniconda.sh
ENV PATH="/opt/conda/bin:$PATH"
# Create and activate Conda environment with Python 3.12, and set it as the default
RUN conda create -n omni python=3.12 && \
echo "source activate omni" > ~/.bashrc
ENV CONDA_DEFAULT_ENV=omni
ENV PATH="/opt/conda/envs/omni/bin:$PATH"
# Set the working directory in the container
WORKDIR /usr/src/app
# Copy project files and requirements
COPY . .
COPY requirements.txt /usr/src/app/requirements.txt
# Initialize Git LFS and pull LFS files
RUN git lfs install && \
git lfs pull
# Install dependencies from requirements.txt with specific opencv-python-headless version
RUN . /opt/conda/etc/profile.d/conda.sh && conda activate omni && \
pip uninstall -y opencv-python opencv-python-headless && \
pip install --no-cache-dir opencv-python-headless==4.8.1.78 && \
pip install -r requirements.txt && \
pip install huggingface_hub
# Run download.py to fetch model weights and convert safetensors to .pt format
# RUN . /opt/conda/etc/profile.d/conda.sh && conda activate omni && \
# python download.py && \
# echo "Contents of weights directory:" && \
# ls -lR weights && \
# python weights/convert_safetensor_to_pt.py
# Expose the default Gradio port
EXPOSE 7861
# Configure Gradio to be accessible externally
ENV GRADIO_SERVER_NAME="0.0.0.0"
# Copy and set permissions for entrypoint script
# COPY entrypoint.sh /usr/src/app/entrypoint.sh
# RUN chmod +x /usr/src/app/entrypoint.sh
# To debug, keep the container running
# CMD ["tail", "-f", "/dev/null"]
################################################################################################
# virtual display related setup --> from anthropic-quickstarts/computer-use-demo/Dockerfile
ENV DEBIAN_FRONTEND=noninteractive
ENV DEBIAN_PRIORITY=high
RUN apt-get update && \
apt-get -y upgrade && \
apt-get -y install \
# UI Requirements
xvfb \
xterm \
xdotool \
scrot \
imagemagick \
sudo \
mutter \
x11vnc \
# Python/pyenv reqs
build-essential \
libssl-dev \
zlib1g-dev \
libbz2-dev \
libreadline-dev \
libsqlite3-dev \
curl \
git \
libncursesw5-dev \
xz-utils \
tk-dev \
libxml2-dev \
libxmlsec1-dev \
libffi-dev \
liblzma-dev \
# Network tools
net-tools \
netcat \
# PPA req
software-properties-common && \
# Userland apps
sudo add-apt-repository ppa:mozillateam/ppa && \
sudo apt-get install -y --no-install-recommends \
libreoffice \
firefox-esr \
x11-apps \
xpdf \
gedit \
xpaint \
tint2 \
galculator \
pcmanfm \
unzip && \
apt-get clean
# Install noVNC
RUN git clone --branch v1.5.0 https://github.com/novnc/noVNC.git /opt/noVNC && \
git clone --branch v0.12.0 https://github.com/novnc/websockify /opt/noVNC/utils/websockify && \
ln -s /opt/noVNC/vnc.html /opt/noVNC/index.html
# setup user
ENV USERNAME=computeruse
ENV HOME=/home/$USERNAME
RUN useradd -m -s /bin/bash -d $HOME $USERNAME
RUN echo "${USERNAME} ALL=(ALL) NOPASSWD: ALL" >> /etc/sudoers
USER computeruse
WORKDIR $HOME
# setup python
RUN git clone https://github.com/pyenv/pyenv.git ~/.pyenv && \
cd ~/.pyenv && src/configure && make -C src && cd .. && \
echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.bashrc && \
echo 'command -v pyenv >/dev/null || export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.bashrc && \
echo 'eval "$(pyenv init -)"' >> ~/.bashrc
ENV PYENV_ROOT="$HOME/.pyenv"
ENV PATH="$PYENV_ROOT/bin:$PATH"
ENV PYENV_VERSION_MAJOR=3
ENV PYENV_VERSION_MINOR=11
ENV PYENV_VERSION_PATCH=6
ENV PYENV_VERSION=$PYENV_VERSION_MAJOR.$PYENV_VERSION_MINOR.$PYENV_VERSION_PATCH
RUN eval "$(pyenv init -)" && \
pyenv install $PYENV_VERSION && \
pyenv global $PYENV_VERSION && \
pyenv rehash
ENV PATH="$HOME/.pyenv/shims:$HOME/.pyenv/bin:$PATH"
RUN python -m pip install --upgrade pip==23.1.2 setuptools==58.0.4 wheel==0.40.0 && \
python -m pip config set global.disable-pip-version-check true
# only reinstall if requirements.txt changes
# COPY --chown=$USERNAME:$USERNAME computer_use_demo/requirements.txt $HOME/computer_use_demo/requirements.txt
# RUN python -m pip install -r $HOME/computer_use_demo/requirements.txt
# setup desktop env & app
# COPY --chown=$USERNAME:$USERNAME image/ $HOME
# COPY --chown=$USERNAME:$USERNAME computer_use_demo/ $HOME/computer_use_demo/
ARG DISPLAY_NUM=1
ARG HEIGHT=768
ARG WIDTH=1024
ENV DISPLAY_NUM=$DISPLAY_NUM
ENV HEIGHT=$HEIGHT
ENV WIDTH=$WIDTH
# Set the entrypoint
# ENTRYPOINT ["/usr/src/app/entrypoint.sh"]
# docker build . -t omniparser-x-demo:local # manually build the docker image (optional)

27
demo.ipynb
View File

@@ -14,7 +14,7 @@
}
],
"source": [
"from utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"from util.utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"import torch\n",
"from ultralytics import YOLO\n",
"from PIL import Image\n",
@@ -48,9 +48,9 @@
"source": [
"# two choices for caption model: fine-tuned blip2 or florence2\n",
"import importlib\n",
"import utils\n",
"importlib.reload(utils)\n",
"from utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"import util.utils\n",
"importlib.reload(util.utils)\n",
"from util.utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"# caption_model_processor = get_caption_model_processor(model_name=\"blip2\", model_name_or_path=\"weights/icon_caption_blip2\", device=device)\n",
"caption_model_processor = get_caption_model_processor(model_name=\"florence2\", model_name_or_path=\"weights/icon_caption_florence\", device=device)\n",
"\n"
@@ -102,9 +102,9 @@
"source": [
"# reload utils\n",
"import importlib\n",
"import utils\n",
"importlib.reload(utils)\n",
"from utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"import util.utils\n",
"importlib.reload(util.utils)\n",
"from util.utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"\n",
"image_path = 'imgs/google_page.png'\n",
"image_path = 'imgs/windows_home.png'\n",
@@ -167,9 +167,9 @@
"# run on cpu!!!\n",
"# reload utils\n",
"import importlib\n",
"import utils\n",
"importlib.reload(utils)\n",
"from utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"import util.utils\n",
"importlib.reload(util.utils)\n",
"from util.utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model\n",
"\n",
"image_path = 'imgs/google_page.png'\n",
"image_path = 'imgs/windows_home.png'\n",
@@ -447,13 +447,6 @@
"source": [
"parsed_content_list[-1]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {

109
demo/gui_agent/llm_utils/llm_utils.py
View File

@@ -1,109 +0,0 @@
import os
import re
import ast
import base64
def is_image_path(text):
# Checking if the input text ends with typical image file extensions
image_extensions = (".jpg", ".jpeg", ".png", ".gif", ".bmp", ".tiff", ".tif")
if text.endswith(image_extensions):
return True
else:
return False
def encode_image(image_path):
"""Encode image file to base64."""
with open(image_path, "rb") as image_file:
return base64.b64encode(image_file.read()).decode("utf-8")
def is_url_or_filepath(input_string):
# Check if input_string is a URL
url_pattern = re.compile(
r"http[s]?://(?:[a-zA-Z]|[0-9]|[$-_@.&+]|[!*\\(\\),]|(?:%[0-9a-fA-F][0-9a-fA-F]))+"
)
if url_pattern.match(input_string):
return "URL"
# Check if input_string is a file path
file_path = os.path.abspath(input_string)
if os.path.exists(file_path):
return "File path"
return "Invalid"
def extract_data(input_string, data_type):
# Regular expression to extract content starting from '```python' until the end if there are no closing backticks
pattern = f"```{data_type}" + r"(.*?)(```|$)"
# Extract content
# re.DOTALL allows '.' to match newlines as well
matches = re.findall(pattern, input_string, re.DOTALL)
# Return the first match if exists, trimming whitespace and ignoring potential closing backticks
return matches[0][0].strip() if matches else input_string
def parse_input(code):
"""Use AST to parse the input string and extract the function name, arguments, and keyword arguments."""
def get_target_names(target):
"""Recursively get all variable names from the assignment target."""
if isinstance(target, ast.Name):
return [target.id]
elif isinstance(target, ast.Tuple):
names = []
for elt in target.elts:
names.extend(get_target_names(elt))
return names
return []
def extract_value(node):
"""提取 AST 节点的实际值"""
if isinstance(node, ast.Constant):
return node.value
elif isinstance(node, ast.Name):
# TODO: a better way to handle variables
raise ValueError(
f"Arguments should be a Constant, got a variable {node.id} instead."
)
# 添加其他需要处理的 AST 节点类型
return None
try:
tree = ast.parse(code)
for node in ast.walk(tree):
if isinstance(node, ast.Assign):
targets = []
for t in node.targets:
targets.extend(get_target_names(t))
if isinstance(node.value, ast.Call):
func_name = node.value.func.id
args = [ast.dump(arg) for arg in node.value.args]
kwargs = {
kw.arg: extract_value(kw.value) for kw in node.value.keywords
}
print(f"Input: {code.strip()}")
print(f"Output Variables: {targets}")
print(f"Function Name: {func_name}")
print(f"Arguments: {args}")
print(f"Keyword Arguments: {kwargs}")
elif isinstance(node, ast.Expr) and isinstance(node.value, ast.Call):
targets = []
func_name = extract_value(node.value.func)
args = [extract_value(arg) for arg in node.value.args]
kwargs = {kw.arg: extract_value(kw.value) for kw in node.value.keywords}
except SyntaxError:
print(f"Input: {code.strip()}")
print("No match found")
return targets, func_name, args, kwargs
if __name__ == "__main__":
import json
s='{"Reasoning": "The Docker icon has been successfully clicked, and the Docker application should now be opening. No further actions are required.", "Next Action": None}'
json_str = json.loads(s)
print(json_str)

117
demo/gui_agent/llm_utils/oai.py
View File

@@ -1,117 +0,0 @@
import os
import logging
import base64
import requests
# from computer_use_demo.gui_agent.llm_utils import is_image_path, encode_image
def is_image_path(text):
# Checking if the input text ends with typical image file extensions
image_extensions = (".jpg", ".jpeg", ".png", ".gif", ".bmp", ".tiff", ".tif")
if text.endswith(image_extensions):
return True
else:
return False
def encode_image(image_path):
"""Encode image file to base64."""
with open(image_path, "rb") as image_file:
return base64.b64encode(image_file.read()).decode("utf-8")
# from openai import OpenAI
# client = OpenAI(
# api_key=os.environ.get("OPENAI_API_KEY")
# )
def run_oai_interleaved(messages: list, system: str, llm: str, api_key: str, max_tokens=256, temperature=0):
api_key = api_key or os.environ.get("OPENAI_API_KEY")
if not api_key:
raise ValueError("OPENAI_API_KEY is not set")
headers = {"Content-Type": "application/json",
"Authorization": f"Bearer {api_key}"}
final_messages = [{"role": "system", "content": system}]
# image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
if type(messages) == list:
for item in messages:
contents = []
if isinstance(item, dict):
for cnt in item["content"]:
if isinstance(cnt, str):
if is_image_path(cnt):
base64_image = encode_image(cnt)
content = {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}
else:
content = {"type": "text", "text": cnt}
else:
# in this case it is a text block from anthropic
content = {"type": "text", "text": str(cnt)}
contents.append(content)
message = {"role": 'user', "content": contents}
else: # str
contents.append({"type": "text", "text": item})
message = {"role": "user", "content": contents}
final_messages.append(message)
elif isinstance(messages, str):
final_messages = [{"role": "user", "content": messages}]
# import pdb; pdb.set_trace()
print("[oai] sending messages:", {"role": "user", "content": messages})
payload = {
"model": llm,
"messages": final_messages,
"max_tokens": max_tokens,
"temperature": temperature,
# "stop": stop,
}
# from IPython.core.debugger import Pdb; Pdb().set_trace()
response = requests.post(
"https://api.openai.com/v1/chat/completions", headers=headers, json=payload
)
try:
text = response.json()['choices'][0]['message']['content']
token_usage = int(response.json()['usage']['total_tokens'])
return text, token_usage
# return error message if the response is not successful
except Exception as e:
print(f"Error in interleaved openAI: {e}. This may due to your invalid OPENAI_API_KEY. Please check the response: {response.json()} ")
return response.json()
if __name__ == "__main__":
api_key = os.environ.get("OPENAI_API_KEY")
if not api_key:
raise ValueError("OPENAI_API_KEY is not set")
text, token_usage = run_oai_interleaved(
messages= [{"content": [
"What is in the screenshot?",
"./tmp/outputs/screenshot_0b04acbb783d4706bc93873d17ba8c05.png"],
"role": "user"
}],
llm="gpt-4o-mini",
system="You are a helpful assistant",
api_key=api_key,
max_tokens=256,
temperature=0)
print(text, token_usage)
# There is an introduction describing the Calyx... 36986

107
demo/gui_agent/llm_utils/qwen.py
View File

@@ -1,107 +0,0 @@
import os
import logging
import base64
import requests
import dashscope
# from computer_use_demo.gui_agent.llm_utils import is_image_path, encode_image
def is_image_path(text):
return False
def encode_image(image_path):
return ""
def run_qwen(messages: list, system: str, llm: str, api_key: str, max_tokens=256, temperature=0):
api_key = api_key or os.environ.get("QWEN_API_KEY")
if not api_key:
raise ValueError("QWEN_API_KEY is not set")
dashscope.api_key = api_key
# from IPython.core.debugger import Pdb; Pdb().set_trace()
final_messages = [{"role": "system", "content": [{"text": system}]}]
# image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
if type(messages) == list:
for item in messages:
contents = []
if isinstance(item, dict):
for cnt in item["content"]:
if isinstance(cnt, str):
if is_image_path(cnt):
# base64_image = encode_image(cnt)
content = [{"image": cnt}]
# content = {"type": "image_url", "image_url": {"url": image_url}}
else:
content = {"text": cnt}
contents.append(content)
message = {"role": item["role"], "content": contents}
else: # str
contents.append({"text": item})
message = {"role": "user", "content": contents}
final_messages.append(message)
print("[qwen-vl] sending messages:", final_messages)
response = dashscope.MultiModalConversation.call(
model='qwen-vl-max-0809',
messages=final_messages
)
# from IPython.core.debugger import Pdb; Pdb().set_trace()
try:
text = response.output.choices[0].message.content[0]['text']
usage = response.usage
if "total_tokens" not in usage:
token_usage = int(usage["input_tokens"] + usage["output_tokens"])
else:
token_usage = int(usage["total_tokens"])
return text, token_usage
# return response.json()['choices'][0]['message']['content']
# return error message if the response is not successful
except Exception as e:
print(f"Error in interleaved openAI: {e}. This may due to your invalid OPENAI_API_KEY. Please check the response: {response.json()} ")
return response.json()
if __name__ == "__main__":
api_key = os.environ.get("QWEN_API_KEY")
if not api_key:
raise ValueError("QWEN_API_KEY is not set")
dashscope.api_key = api_key
final_messages = [{"role": "user",
"content": [
{"text": "What is in the screenshot?"},
{"image": "./tmp/outputs/screenshot_0b04acbb783d4706bc93873d17ba8c05.png"}
]
}
]
response = dashscope.MultiModalConversation.call(model='qwen-vl-max-0809', messages=final_messages)
print(response)
text = response.output.choices[0].message.content[0]['text']
usage = response.usage
if "total_tokens" not in usage:
if "image_tokens" in usage:
token_usage = usage["input_tokens"] + usage["output_tokens"] + usage["image_tokens"]
else:
token_usage = usage["input_tokens"] + usage["output_tokens"]
else:
token_usage = usage["total_tokens"]
print(text, token_usage)
# The screenshot is from a video game... 1387

44
demo/gui_agent/llm_utils/run_llm.py
View File

@@ -1,44 +0,0 @@
import base64
import logging
from .oai import run_oai_interleaved
from .gemini import run_gemini_interleaved
def run_llm(prompt, llm="gpt-4o-mini", max_tokens=256, temperature=0, stop=None):
log_prompt(prompt)
# turn string prompt into list
if isinstance(prompt, str):
prompt = [prompt]
elif isinstance(prompt, list):
pass
else:
raise ValueError(f"Invalid prompt type: {type(prompt)}")
if llm.startswith("gpt"): # gpt series
out = run_oai_interleaved(
prompt,
llm,
max_tokens,
temperature,
stop
)
elif llm.startswith("gemini"): # gemini series
out = run_gemini_interleaved(
prompt,
llm,
max_tokens,
temperature,
stop
)
else:
raise ValueError(f"Invalid llm: {llm}")
logging.info(
f"========Output for {llm}=======\n{out}\n============================")
return out
def log_prompt(prompt):
prompt_display = [prompt] if isinstance(prompt, str) else prompt
prompt_display = "\n\n".join(prompt_display)
logging.info(
f"========Prompt=======\n{prompt_display}\n============================")

236
demo/loop.py
View File

@@ -1,236 +0,0 @@
"""
Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
"""
import time
import json
import asyncio
import platform
from collections.abc import Callable
from datetime import datetime
from enum import StrEnum
from typing import Any, cast, Dict
from anthropic import Anthropic, AnthropicBedrock, AnthropicVertex, APIResponse
from anthropic.types import (
ToolResultBlockParam,
TextBlock,
)
from anthropic.types.beta import (
BetaContentBlock,
BetaContentBlockParam,
BetaImageBlockParam,
BetaMessage,
BetaMessageParam,
BetaTextBlockParam,
BetaToolResultBlockParam,
)
from tools import BashTool, ComputerTool, EditTool, ToolCollection, ToolResult
import torch
from gui_agent.anthropic_agent import AnthropicActor
from executor.anthropic_executor import AnthropicExecutor
from omniparser_agent.vlm_agent import OmniParser, VLMAgent
from tools.colorful_text import colorful_text_showui, colorful_text_vlm
from tools.screen_capture import get_screenshot
from gui_agent.llm_utils.oai import encode_image
BETA_FLAG = "computer-use-2024-10-22"
class APIProvider(StrEnum):
ANTHROPIC = "anthropic"
BEDROCK = "bedrock"
VERTEX = "vertex"
OPENAI = "openai"
QWEN = "qwen"
PROVIDER_TO_DEFAULT_MODEL_NAME: dict[APIProvider, str] = {
APIProvider.ANTHROPIC: "claude-3-5-sonnet-20241022",
APIProvider.BEDROCK: "anthropic.claude-3-5-sonnet-20241022-v2:0",
APIProvider.VERTEX: "claude-3-5-sonnet-v2@20241022",
# APIProvider.OPENAI: "gpt-4o",
# APIProvider.QWEN: "qwen2vl",
}
# This system prompt is optimized for the Docker environment in this repository and
# specific tool combinations enabled.
# We encourage modifying this system prompt to ensure the model has context for the
# environment it is running in, and to provide any additional information that may be
# helpful for the task at hand.
SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
* You are utilizing a Windows system with internet access.
* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
</SYSTEM_CAPABILITY>
"""
import base64
from PIL import Image
from io import BytesIO
def sampling_loop_sync(
*,
model: str,
provider: APIProvider | None,
system_prompt_suffix: str,
messages: list[BetaMessageParam],
output_callback: Callable[[BetaContentBlock], None],
tool_output_callback: Callable[[ToolResult, str], None],
api_response_callback: Callable[[APIResponse[BetaMessage]], None],
api_key: str,
only_n_most_recent_images: int | None = 2,
max_tokens: int = 4096,
selected_screen: int = 0
):
"""
Synchronous agentic sampling loop for the assistant/tool interaction of computer use.
"""
print('in sampling_loop_sync, model:', model)
if model == "claude-3-5-sonnet-20241022":
omniparser = OmniParser(url="http://127.0.0.1:8000/send_text/",
selected_screen=selected_screen,)
# Register Actor and Executor
actor = AnthropicActor(
model=model,
provider=provider,
system_prompt_suffix=system_prompt_suffix,
api_key=api_key,
api_response_callback=api_response_callback,
max_tokens=max_tokens,
only_n_most_recent_images=only_n_most_recent_images,
selected_screen=selected_screen
)
# from IPython.core.debugger import Pdb; Pdb().set_trace()
executor = AnthropicExecutor(
output_callback=output_callback,
tool_output_callback=tool_output_callback,
selected_screen=selected_screen
)
elif model == "omniparser + gpt-4o" or model == "omniparser + phi35v":
omniparser = OmniParser(url="http://127.0.0.1:8000/send_text/",
selected_screen=selected_screen,)
actor = VLMAgent(
model=model,
provider=provider,
system_prompt_suffix=system_prompt_suffix,
api_key=api_key,
api_response_callback=api_response_callback,
selected_screen=selected_screen,
output_callback=output_callback,
)
executor = AnthropicExecutor(
output_callback=output_callback,
tool_output_callback=tool_output_callback,
selected_screen=selected_screen
)
# elif model == "gpt-4o + ShowUI" or model == "qwen2vl + ShowUI":
# planner = VLMPlanner(
# model=model,
# provider=provider,
# system_prompt_suffix=system_prompt_suffix,
# api_key=api_key,
# api_response_callback=api_response_callback,
# selected_screen=selected_screen,
# output_callback=output_callback,
# )
# if torch.cuda.is_available(): device = torch.device("cuda")
# elif torch.backends.mps.is_available(): device = torch.device("mps")
# else: device = torch.device("cpu") # support: 'cpu', 'mps', 'cuda'
# print(f"showUI-2B inited on device: {device}.")
# actor = ShowUIActor(
# model_path="./showui-2b/",
# # Replace with your local path, e.g., "C:\\code\\ShowUI-2B", "/Users/your_username/ShowUI-2B/".
# device=device,
# split='web', # 'web' or 'phone'
# selected_screen=selected_screen,
# output_callback=output_callback,
# )
# executor = ShowUIExecutor(
# output_callback=output_callback,
# tool_output_callback=tool_output_callback,
# selected_screen=selected_screen
# )
else:
raise ValueError(f"Model {model} not supported")
print(f"Model Inited: {model}, Provider: {provider}")
tool_result_content = None
print(f"Start the message loop. User messages: {messages}")
if model == "claude-3-5-sonnet-20241022": # Anthropic loop
while True:
parsed_screen = omniparser() # parsed_screen: {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, "screen_info"}
import pdb; pdb.set_trace()
screen_info_block = TextBlock(text='Below is the structured accessibility information of the current UI screen, which includes text and icons you can operate on, take these information into account when you are making the prediction for the next action. Note you will still need to take screenshot to get the image: \n' + parsed_screen['screen_info'], type='text')
# # messages[-1]['content'].append(screen_info_block)
screen_info_dict = {"role": "user", "content": [screen_info_block]}
messages.append(screen_info_dict)
response = actor(messages=messages)
for message, tool_result_content in executor(response, messages):
yield message
if not tool_result_content:
return messages
messages.append({"content": tool_result_content, "role": "user"})
elif model == "omniparser + gpt-4o" or model == "omniparser + phi35v":
while True:
parsed_screen = omniparser()
response, vlm_response_json = actor(messages=messages, parsed_screen=parsed_screen)
for message, tool_result_content in executor(response, messages):
yield message
if not tool_result_content:
return messages
# import pdb; pdb.set_trace()
# messages.append({"role": "user",
# "content": ["History plan:\n" + str(vlm_response_json['Reasoning'])]})
# messages.append({"content": tool_result_content, "role": "user"})
elif model == "gpt-4o + ShowUI" or model == "qwen2vl + ShowUI": # ShowUI loop
while True:
vlm_response = planner(messages=messages)
next_action = json.loads(vlm_response).get("Next Action")
yield next_action
if next_action == None or next_action == "" or next_action == "None":
final_sc, final_sc_path = get_screenshot(selected_screen=selected_screen)
output_callback(f'No more actions from {colorful_text_vlm}. End of task. Final State:\n<img src="data:image/png;base64,{encode_image(str(final_sc_path))}">',
sender="bot")
yield None
output_callback(f"{colorful_text_vlm} sending action to {colorful_text_showui}:\n{next_action}", sender="bot")
actor_response = actor(messages=next_action)
yield actor_response
for message, tool_result_content in executor(actor_response, messages):
time.sleep(1)
yield message
# since showui executor has no feedback for now, we use "actor_response" to represent its response
# update messages for the next loop
messages.append({"role": "user",
"content": ["History plan:\n" + str(json.loads(vlm_response)) +
"\nHistory actions:\n" + str(actor_response["content"])]})
print(f"End of loop. Messages: {str(messages)[:100000]}. Total cost: $USD{planner.total_cost:.5f}")

417
demo/omniparser_agent/vlm_agent.py
View File

@@ -1,417 +0,0 @@
import json
import asyncio
import platform
from collections.abc import Callable
from datetime import datetime
from enum import StrEnum
from typing import Any, cast, Dict, Callable
import uuid
import requests
from PIL import Image, ImageDraw
import base64
from io import BytesIO
from anthropic import Anthropic, AnthropicBedrock, AnthropicVertex, APIResponse
from anthropic.types import TextBlock, ToolResultBlockParam
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock, BetaMessageParam, BetaUsage
from tools.screen_capture import get_screenshot
from gui_agent.llm_utils.oai import run_oai_interleaved, encode_image
from gui_agent.llm_utils.qwen import run_qwen
from gui_agent.llm_utils.llm_utils import extract_data
from tools.colorful_text import colorful_text_showui, colorful_text_vlm
SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
* You are utilizing a Windows system with internet access.
* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
</SYSTEM_CAPABILITY>
"""
class OmniParser:
def __init__(self,
url: str,
selected_screen: int = 0) -> None:
self.url = url
self.selected_screen = selected_screen
def __call__(self,):
screenshot, screenshot_path = get_screenshot(selected_screen=self.selected_screen)
screenshot_path = str(screenshot_path)
image_base64 = encode_image(screenshot_path)
# response = requests.post(self.url, json={"base64_image": image_base64, 'prompt': 'omniparser process'})
# response_json = response.json()
# example response_json: {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, "latency": 0.1}
# Debug
response_json = {"som_image_base64": image_base64, "parsed_content_list": ['debug1', 'debug2'], "latency": 0.1}
print('omniparser latency:', response_json['latency'])
response_json = self.reformat_messages(response_json)
return response_json
def reformat_messages(self, response_json: dict):
parsed_content_list = response_json["parsed_content_list"]
screen_info = ""
# Debug
# for idx, element in enumerate(parsed_content_list):
# element['idx'] = idx
# if element['type'] == 'text':
# # screen_info += f'''<p id={idx} class="text" alt="{element['content']}"> </p>\n'''
# screen_info += f'ID: {idx}, Text: {element["content"]}\n'
# elif element['type'] == 'icon':
# # screen_info += f'''<img id={idx} class="icon" alt="{element['content']}"> </img>\n'''
# screen_info += f'ID: {idx}, Icon: {element["content"]}\n'
response_json['screen_info'] = screen_info
return response_json
class VLMAgent:
def __init__(
self,
model: str,
provider: str,
system_prompt_suffix: str,
api_key: str,
output_callback: Callable,
api_response_callback: Callable,
max_tokens: int = 4096,
only_n_most_recent_images: int | None = None,
selected_screen: int = 0,
print_usage: bool = True,
):
if model == "gpt-4o + ShowUI":
self.model = "gpt-4o-2024-11-20"
elif model == "gpt-4o-mini + ShowUI":
self.model = "gpt-4o-mini" # "gpt-4o-mini"
elif model == "qwen2vl + ShowUI":
self.model = "qwen2vl"
elif model == "omniparser + gpt-4o":
self.model = "gpt-4o-2024-11-20"
else:
raise ValueError(f"Model {model} not supported")
self.provider = provider
self.system_prompt_suffix = system_prompt_suffix
self.api_key = api_key
self.api_response_callback = api_response_callback
self.max_tokens = max_tokens
self.only_n_most_recent_images = only_n_most_recent_images
self.selected_screen = selected_screen
self.output_callback = output_callback
self.print_usage = print_usage
self.total_token_usage = 0
self.total_cost = 0
self.system = (
# f"{SYSTEM_PROMPT}{' ' + system_prompt_suffix if system_prompt_suffix else ''}"
f"{system_prompt_suffix}"
)
def __call__(self, messages: list, parsed_screen: list[str, list]):
# example parsed_screen: {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, "screen_info"}
screen_info = parsed_screen["screen_info"]
# drop looping actions msg, byte image etc
planner_messages = messages
# planner_messages = _message_filter_callback(messages)
print(f"filtered_messages: {planner_messages}\n\n", "full messages:", messages)
# import pdb; pdb.set_trace()
planner_messages = _keep_latest_images(planner_messages)
# if self.only_n_most_recent_images:
# _maybe_filter_to_n_most_recent_images(planner_messages, self.only_n_most_recent_images)
system = self._get_system_prompt(screen_info) + self.system_prompt_suffix
# Take a screenshot
screenshot, screenshot_path = get_screenshot(selected_screen=self.selected_screen)
screen_width, screen_height = screenshot.size
screenshot_path = str(screenshot_path)
image_base64 = encode_image(screenshot_path)
som_image_data = base64.b64decode(parsed_screen['som_image_base64'])
som_screenshot_path = f"./tmp/outputs/screenshot_som_{uuid.uuid4().hex}.png"
with open(som_screenshot_path, "wb") as f:
f.write(som_image_data)
self.output_callback(f'Screenshot for {colorful_text_vlm}:\n<img src="data:image/png;base64,{image_base64}">',
sender="bot")
self.output_callback(f'Set of Marks Screenshot for {colorful_text_vlm}:\n<img src="data:image/png;base64,{parsed_screen['som_image_base64']}">', sender="bot")
if isinstance(planner_messages[-1], dict):
if not isinstance(planner_messages[-1]["content"], list):
planner_messages[-1]["content"] = [planner_messages[-1]["content"]]
planner_messages[-1]["content"].append(screenshot_path)
planner_messages[-1]["content"].append(som_screenshot_path)
print(f"Sending messages to VLMPlanner : {planner_messages}")
if "gpt" in self.model:
vlm_response, token_usage = run_oai_interleaved(
messages=planner_messages,
system=system,
llm=self.model,
api_key=self.api_key,
max_tokens=self.max_tokens,
temperature=0,
)
print(f"oai token usage: {token_usage}")
self.total_token_usage += token_usage
self.total_cost += (token_usage * 0.15 / 1000000) # https://openai.com/api/pricing/
elif "qwen" in self.model:
vlm_response, token_usage = run_qwen(
messages=planner_messages,
system=system,
llm=self.model,
api_key=self.api_key,
max_tokens=self.max_tokens,
temperature=0,
)
print(f"qwen token usage: {token_usage}")
self.total_token_usage += token_usage
self.total_cost += (token_usage * 0.02 / 7.25 / 1000) # 1USD=7.25CNY, https://help.aliyun.com/zh/dashscope/developer-reference/tongyi-qianwen-vl-plus-api
elif "phi" in self.model:
pass # TODO
else:
raise ValueError(f"Model {self.model} not supported")
print(f"VLMPlanner response: {vlm_response}")
if self.print_usage:
print(f"VLMPlanner total token usage so far: {self.total_token_usage}. Total cost so far: $USD{self.total_cost:.5f}")
vlm_response_json = extract_data(vlm_response, "json")
vlm_response_json = json.loads(vlm_response_json)
# map "box_id" to "idx" in parsed_screen, and output the xy coordinate of bbox
# TODO add try except for the case when "box_id" is not in the response
# if 'Box ID' in vlm_response_json:
try:
bbox = parsed_screen["parsed_content_list"][int(vlm_response_json["Box ID"])]["bbox"]
vlm_response_json["coordinate"] = [int((bbox[0] + bbox[2]) / 2 * screen_width), int((bbox[1] + bbox[3]) / 2 * screen_height)]
# draw a circle on the screenshot image to indicate the action
self.draw_action(vlm_response_json, image_base64)
except:
print("No Box ID in the response.")
# vlm_plan_str = '\n'.join([f'{key}: {value}' for key, value in json.loads(response).items()])
vlm_plan_str = ""
for key, value in vlm_response_json.items():
if key == "Reasoning":
vlm_plan_str += f'{value}'
else:
vlm_plan_str += f'\n{key}: {value}'
# self.output_callback(f"{colorful_text_vlm}:\n{vlm_plan_str}", sender="bot")
# construct the response so that anthropicExcutor can execute the tool
analysis = BetaTextBlock(text=vlm_plan_str, type='text')
if 'coordinate' in vlm_response_json:
move_cursor_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
input={'action': 'mouse_move', 'coordinate': vlm_response_json["coordinate"]},
name='computer', type='tool_use')
response_content = [analysis, move_cursor_block]
else:
response_content = [analysis]
if vlm_response_json["Next Action"] == "type":
click_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}', input={'action': 'left_click'}, name='computer', type='tool_use')
sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
input={'action': vlm_response_json["Next Action"], 'text': vlm_response_json["value"]},
name='computer', type='tool_use')
response_content.extend([click_block, sim_content_block])
elif vlm_response_json["Next Action"] == "None":
print("Task paused/completed.")
else:
sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
input={'action': vlm_response_json["Next Action"]},
name='computer', type='tool_use')
response_content.append(sim_content_block)
response = BetaMessage(id=f'toolu_{uuid.uuid4()}', content=response_content, model='', role='assistant', type='message', stop_reason='tool_use', usage=BetaUsage(input_tokens=0, output_tokens=0))
return response, vlm_response_json
def _api_response_callback(self, response: APIResponse):
self.api_response_callback(response)
def reformat_messages(self, messages: list):
pass
def _get_system_prompt(self, screen_info: str = ""):
datetime_str = datetime.now().strftime("%A, %B %d, %Y")
os_name = platform.system()
return f"""
You are using an {os_name} device.
You are able to use a mouse and keyboard to interact with the computer based on the given task and screenshot.
You can only interact with the desktop GUI (no terminal or application menu access).
You may be given some history plan and actions, this is the response from the previous loop.
You should carefully consider your plan base on the task, screenshot, and history actions.
Here is the list of all detected bounding boxes by IDs on the screen and their description:{screen_info}
Your available "Next Action" only include:
- type: type a string of text.
- left_click: Describe the ui element to be clicked.
- enter: Press an enter key.
- escape: Press an ESCAPE key.
- hover: Describe the ui element to be hovered.
- scroll: Scroll the screen, you must specify up or down.
- press: Describe the ui element to be pressed.
Based on the visual information from the screenshot image and the detected bounding boxes, please determine the next action, the Box ID you should operate on, and the value (if the action is 'type') in order to complete the task.
Output format:
```json
{{
"Reasoning": str, # describe what is in the current screen, taking into account the history, then describe your step-by-step thoughts on how to achieve the task, choose one action from available actions at a time.
"Next Action": "action_type, action description" | "None" # one action at a time, describe it in short and precisely.
'Box ID': n,
'value': "xxx" # if the action is type, you should provide the text to type.
}}
```
One Example:
```json
{{
"Reasoning": "The current screen shows google result of amazon, in previous action I have searched amazon on google. Then I need to click on the first search results to go to amazon.com.",
"Next Action": "left_click",
'Box ID': m,
}}
```
Another Example:
```json
{{
"Reasoning": "The current screen shows the front page of amazon. There is no previous action. Therefore I need to type "Apple watch" in the search bar.",
"Next Action": "type",
'Box ID': n,
'value': "Apple watch"
}}
```
IMPORTANT NOTES:
1. You should only give a single action at a time.
2. You should give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task.
3. Attach the next action prediction in the "Next Action".
4. You should not include other actions, such as keyboard shortcuts.
5. When the task is completed, you should say "Next Action": "None" in the json field.
"""
def draw_action(self, vlm_response_json, image_base64):
# draw a circle using the coordinate in parsed_screen['som_image_base64']
image_data = base64.b64decode(image_base64)
image = Image.open(BytesIO(image_data))
draw = ImageDraw.Draw(image)
x, y = vlm_response_json["coordinate"]
radius = 10
draw.ellipse((x - radius, y - radius, x + radius, y + radius), outline='red', width=3)
buffered = BytesIO()
image.save('demo.png')
image.save(buffered, format="PNG")
image_with_circle_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
self.output_callback(f'Action performed on the Screenshot (red circle), for {colorful_text_vlm}:\n<img src="data:image/png;base64,{image_with_circle_base64}">', sender="bot")
def _keep_latest_images(messages):
for i in range(len(messages)-1):
if isinstance(messages[i]["content"], list):
for cnt in messages[i]["content"]:
if isinstance(cnt, str):
if cnt.endswith((".jpg", ".jpeg", ".png", ".gif", ".bmp", ".tiff", ".tif")):
messages[i]["content"].remove(cnt)
return messages
def _maybe_filter_to_n_most_recent_images(
messages: list[BetaMessageParam],
images_to_keep: int,
min_removal_threshold: int = 10,
):
"""
With the assumption that images are screenshots that are of diminishing value as
the conversation progresses, remove all but the final `images_to_keep` tool_result
images in place, with a chunk of min_removal_threshold to reduce the amount we
break the implicit prompt cache.
"""
if images_to_keep is None:
return messages
tool_result_blocks = cast(
list[ToolResultBlockParam],
[
item
for message in messages
for item in (
message["content"] if isinstance(message["content"], list) else []
)
if isinstance(item, dict) and item.get("type") == "tool_result"
],
)
total_images = sum(
1
for tool_result in tool_result_blocks
for content in tool_result.get("content", [])
if isinstance(content, dict) and content.get("type") == "image"
)
images_to_remove = total_images - images_to_keep
# for better cache behavior, we want to remove in chunks
images_to_remove -= images_to_remove % min_removal_threshold
for tool_result in tool_result_blocks:
if isinstance(tool_result.get("content"), list):
new_content = []
for content in tool_result.get("content", []):
if isinstance(content, dict) and content.get("type") == "image":
if images_to_remove > 0:
images_to_remove -= 1
continue
new_content.append(content)
tool_result["content"] = new_content
def _message_filter_callback(messages):
filtered_list = []
try:
for msg in messages:
if msg.get('role') in ['user']:
if not isinstance(msg["content"], list):
msg["content"] = [msg["content"]]
if isinstance(msg["content"][0], TextBlock):
filtered_list.append(str(msg["content"][0].text)) # User message
elif isinstance(msg["content"][0], str):
filtered_list.append(msg["content"][0]) # User message
else:
print("[_message_filter_callback]: drop message", msg)
continue
# elif msg.get('role') in ['assistant']:
# if isinstance(msg["content"][0], TextBlock):
# msg["content"][0] = str(msg["content"][0].text)
# elif isinstance(msg["content"][0], BetaTextBlock):
# msg["content"][0] = str(msg["content"][0].text)
# elif isinstance(msg["content"][0], BetaToolUseBlock):
# msg["content"][0] = str(msg['content'][0].input)
# elif isinstance(msg["content"][0], Dict) and msg["content"][0]["content"][-1]["type"] == "image":
# msg["content"][0] = f'<img src="data:image/png;base64,{msg["content"][0]["content"][-1]["source"]["data"]}">'
# else:
# print("[_message_filter_callback]: drop message", msg)
# continue
# filtered_list.append(msg["content"][0]) # User message
else:
print("[_message_filter_callback]: drop message", msg)
continue
except Exception as e:
print("[_message_filter_callback]: error", e)
return filtered_list

78
demo/remote_request.py
View File

@@ -1,78 +0,0 @@
# uvicorn remote_request:app --host 0.0.0.0 --port 8000 --reload
import sys
import os
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model
import torch
from PIL import Image
from typing import Dict, Tuple, List
import base64
config = {
'som_model_path': '../weights/icon_detect_v1_5/model_v1_5.pt',
'device': 'cpu',
'caption_model_name': 'florence2',
'caption_model_path': '../weights/icon_caption_florence',
'BOX_TRESHOLD': 0.05
}
class Omniparser(object):
def __init__(self, config: Dict):
self.config = config
device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.som_model = get_yolo_model(model_path=config['som_model_path'])
self.caption_model_processor = get_caption_model_processor(model_name=config['caption_model_name'], model_name_or_path=config['caption_model_path'], device=device)
print('Omniparser initialized!!!')
def parse(self, image_base64: str):
image_path = '../imgs/demo_image.jpg'
with open(image_path, "wb") as fh:
fh.write(base64.b64decode(image_base64))
print('Parsing image:', image_path)
image = Image.open(image_path)
print('image size:', image.size)
box_overlay_ratio = max(image.size) / 3200
draw_bbox_config = {
'text_scale': 0.8 * box_overlay_ratio,
'text_thickness': max(int(2 * box_overlay_ratio), 1),
'text_padding': max(int(3 * box_overlay_ratio), 1),
'thickness': max(int(3 * box_overlay_ratio), 1),
}
BOX_TRESHOLD = config['BOX_TRESHOLD']
ocr_bbox_rslt, is_goal_filtered = check_ocr_box(image_path, display_img = False, output_bb_format='xyxy', goal_filtering=None, easyocr_args={'paragraph': False, 'text_threshold':0.8}, use_paddleocr=True)
text, ocr_bbox = ocr_bbox_rslt
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(image_path, self.som_model, BOX_TRESHOLD = BOX_TRESHOLD, output_coord_in_ratio=True, ocr_bbox=ocr_bbox,draw_bbox_config=draw_bbox_config, caption_model_processor=self.caption_model_processor, ocr_text=text,use_local_semantics=True, iou_threshold=0.7, scale_img=False, batch_size=128)
with open('../imgs/demo_image_som.jpg', "wb") as fh:
fh.write(base64.b64decode(dino_labled_img))
return dino_labled_img, parsed_content_list
from fastapi import FastAPI
from pydantic import BaseModel
app = FastAPI()
class Item(BaseModel):
base64_image: str
prompt: str
Omniparser = Omniparser(config)
@app.post("/send_text/")
async def send_text(item: Item):
print('start parsing...')
import time
start = time.time()
dino_labled_img, parsed_content_list = Omniparser.parse(item.base64_image)
latency = time.time() - start
print('time:', latency)
return {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, 'latency': latency}

BIN
demo/tmp/outputs/screenshot.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 1.4 MiB

136
demo/tools/bash.py
View File

@@ -1,136 +0,0 @@
import asyncio
import os
from typing import ClassVar, Literal
from anthropic.types.beta import BetaToolBash20241022Param
from .base import BaseAnthropicTool, CLIResult, ToolError, ToolResult
class _BashSession:
"""A session of a bash shell."""
_started: bool
_process: asyncio.subprocess.Process
command: str = "/bin/bash"
_output_delay: float = 0.2 # seconds
_timeout: float = 120.0 # seconds
_sentinel: str = "<<exit>>"
def __init__(self):
self._started = False
self._timed_out = False
async def start(self):
if self._started:
return
self._process = await asyncio.create_subprocess_shell(
self.command,
shell=False,
stdin=asyncio.subprocess.PIPE,
stdout=asyncio.subprocess.PIPE,
stderr=asyncio.subprocess.PIPE,
)
self._started = True
def stop(self):
"""Terminate the bash shell."""
if not self._started:
raise ToolError("Session has not started.")
if self._process.returncode is not None:
return
self._process.terminate()
async def run(self, command: str):
"""Execute a command in the bash shell."""
if not self._started:
raise ToolError("Session has not started.")
if self._process.returncode is not None:
return ToolResult(
system="tool must be restarted",
error=f"bash has exited with returncode {self._process.returncode}",
)
if self._timed_out:
raise ToolError(
f"timed out: bash has not returned in {self._timeout} seconds and must be restarted",
)
# we know these are not None because we created the process with PIPEs
assert self._process.stdin
assert self._process.stdout
assert self._process.stderr
# send command to the process
self._process.stdin.write(
command.encode() + f"; echo '{self._sentinel}'\n".encode()
)
await self._process.stdin.drain()
# read output from the process, until the sentinel is found
output = ""
try:
async with asyncio.timeout(self._timeout):
while True:
await asyncio.sleep(self._output_delay)
data = await self._process.stdout.readline()
if not data:
break
line = data.decode()
output += line
if self._sentinel in line:
output = output.replace(self._sentinel, "")
break
except asyncio.TimeoutError:
self._timed_out = True
raise ToolError(
f"timed out: bash has not returned in {self._timeout} seconds and must be restarted",
) from None
error = await self._process.stderr.read()
error = error.decode()
return CLIResult(output=output.strip(), error=error.strip())
class BashTool(BaseAnthropicTool):
"""
A tool that allows the agent to run bash commands.
The tool parameters are defined by Anthropic and are not editable.
"""
_session: _BashSession | None
name: ClassVar[Literal["bash"]] = "bash"
api_type: ClassVar[Literal["bash_20241022"]] = "bash_20241022"
def __init__(self):
self._session = None
super().__init__()
async def __call__(
self, command: str | None = None, restart: bool = False, **kwargs
):
if restart:
if self._session:
self._session.stop()
self._session = _BashSession()
await self._session.start()
return ToolResult(system="tool has been restarted.")
if self._session is None:
self._session = _BashSession()
await self._session.start()
if command is not None:
return await self._session.run(command)
raise ToolError("no command provided.")
def to_params(self) -> BetaToolBash20241022Param:
return {
"type": self.api_type,
"name": self.name,
}

27
demo/tools/colorful_text.py
View File

@@ -1,27 +0,0 @@
"""
Define some colorful stuffs for better visualization in the chat.
"""
# Define the RGB colors for each letter
colors = {
'S': 'rgb(106, 158, 210)',
'h': 'rgb(111, 163, 82)',
'o': 'rgb(209, 100, 94)',
'w': 'rgb(238, 171, 106)',
'U': 'rgb(0, 0, 0)',
'I': 'rgb(0, 0, 0)',
}
# Construct the colorful "ShowUI" word
colorful_text_showui = "**"+''.join(
f'<span style="color:{colors.get(letter, "black")}">{letter}</span>'
for letter in "ShowUI"
)+"**"
colorful_text_vlm = "**OmniParser Agent**"
colorful_text_user = "**User**"
# print(f"colorful_text_showui: {colorful_text_showui}")
# **<span style="color:rgb(106, 158, 210)">S</span><span style="color:rgb(111, 163, 82)">h</span><span style="color:rgb(209, 100, 94)">o</span><span style="color:rgb(238, 171, 106)">w</span><span style="color:rgb(0, 0, 0)">U</span><span style="color:rgb(0, 0, 0)">I</span>**

519
demo/tools/computer.py
View File

@@ -1,519 +0,0 @@
import subprocess
import platform
import pyautogui
import asyncio
import base64
import os
import time
if platform.system() == "Darwin":
import Quartz # uncomment this line if you are on macOS
from enum import StrEnum
from pathlib import Path
from typing import Literal, TypedDict
from uuid import uuid4
from screeninfo import get_monitors
from PIL import ImageGrab, Image
from functools import partial
from anthropic.types.beta import BetaToolComputerUse20241022Param
from .base import BaseAnthropicTool, ToolError, ToolResult
from .run import run
OUTPUT_DIR = "./tmp/outputs"
TYPING_DELAY_MS = 12
TYPING_GROUP_SIZE = 50
Action = Literal[
"key",
"type",
"mouse_move",
"left_click",
"left_click_drag",
"right_click",
"middle_click",
"double_click",
"screenshot",
"cursor_position",
]
class Resolution(TypedDict):
width: int
height: int
MAX_SCALING_TARGETS: dict[str, Resolution] = {
"XGA": Resolution(width=1024, height=768), # 4:3
"WXGA": Resolution(width=1280, height=800), # 16:10
"FWXGA": Resolution(width=1366, height=768), # ~16:9
}
class ScalingSource(StrEnum):
COMPUTER = "computer"
API = "api"
class ComputerToolOptions(TypedDict):
display_height_px: int
display_width_px: int
display_number: int | None
def chunks(s: str, chunk_size: int) -> list[str]:
return [s[i : i + chunk_size] for i in range(0, len(s), chunk_size)]
def get_screen_details():
screens = get_monitors()
screen_details = []
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s.x)
# Loop through sorted screens and assign positions
primary_index = 0
for i, screen in enumerate(sorted_screens):
if i == 0:
layout = "Left"
elif i == len(sorted_screens) - 1:
layout = "Right"
else:
layout = "Center"
if screen.is_primary:
position = "Primary"
primary_index = i
else:
position = "Secondary"
screen_info = f"Screen {i + 1}: {screen.width}x{screen.height}, {layout}, {position}"
screen_details.append(screen_info)
return screen_details, primary_index
class ComputerTool(BaseAnthropicTool):
"""
A tool that allows the agent to interact with the screen, keyboard, and mouse of the current computer.
Adapted for Windows using 'pyautogui'.
"""
name: Literal["computer"] = "computer"
api_type: Literal["computer_20241022"] = "computer_20241022"
width: int
height: int
display_num: int | None
_screenshot_delay = 2.0
_scaling_enabled = True
@property
def options(self) -> ComputerToolOptions:
width, height = self.scale_coordinates(
ScalingSource.COMPUTER, self.width, self.height
)
return {
"display_width_px": width,
"display_height_px": height,
"display_number": self.display_num,
}
def to_params(self) -> BetaToolComputerUse20241022Param:
return {"name": self.name, "type": self.api_type, **self.options}
def __init__(self, selected_screen: int = 0, is_scaling: bool = False):
super().__init__()
# Get screen width and height using Windows command
self.display_num = None
self.offset_x = 0
self.offset_y = 0
self.selected_screen = selected_screen
self.is_scaling = is_scaling
self.width, self.height = self.get_screen_size()
# Path to cliclick
self.cliclick = "cliclick"
self.key_conversion = {"Page_Down": "pagedown",
"Page_Up": "pageup",
"Super_L": "win",
"Escape": "esc"}
system = platform.system() # Detect platform
if system == "Windows":
screens = get_monitors()
sorted_screens = sorted(screens, key=lambda s: s.x)
if self.selected_screen < 0 or self.selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
screen = sorted_screens[self.selected_screen]
bbox = (screen.x, screen.y, screen.x + screen.width, screen.y + screen.height)
elif system == "Darwin": # macOS
max_displays = 32 # Maximum number of displays to handle
active_displays = Quartz.CGGetActiveDisplayList(max_displays, None, None)[1]
screens = []
for display_id in active_displays:
bounds = Quartz.CGDisplayBounds(display_id)
screens.append({
'id': display_id, 'x': int(bounds.origin.x), 'y': int(bounds.origin.y),
'width': int(bounds.size.width), 'height': int(bounds.size.height),
'is_primary': Quartz.CGDisplayIsMain(display_id) # Check if this is the primary display
})
sorted_screens = sorted(screens, key=lambda s: s['x'])
if self.selected_screen < 0 or self.selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
screen = sorted_screens[self.selected_screen]
bbox = (screen['x'], screen['y'], screen['x'] + screen['width'], screen['y'] + screen['height'])
else: # Linux or other OS
cmd = "xrandr | grep ' primary' | awk '{print $4}'"
try:
# output = subprocess.check_output(cmd, shell=True).decode()
# resolution = output.strip().split()[0]
# width, height = map(int, resolution.split('x'))
# bbox = (0, 0, width, height) # Assuming single primary screen for simplicity
screen = get_monitors()[0]
bbox = (screen.x, screen.y, screen.x + screen.width, screen.y + screen.height)
except subprocess.CalledProcessError:
raise RuntimeError("Failed to get screen resolution on Linux.")
self.offset_x = screen['x'] if system == "Darwin" else screen.x
self.offset_y = screen['y'] if system == "Darwin" else screen.y
self.bbox = bbox
async def __call__(
self,
*,
action: Action,
text: str | None = None,
coordinate: tuple[int, int] | None = None,
**kwargs,
):
print(f"action: {action}, text: {text}, coordinate: {coordinate}, is_scaling: {self.is_scaling}")
if action in ("mouse_move", "left_click_drag"):
if coordinate is None:
raise ToolError(f"coordinate is required for {action}")
if text is not None:
raise ToolError(f"text is not accepted for {action}")
if not isinstance(coordinate, (list, tuple)) or len(coordinate) != 2:
raise ToolError(f"{coordinate} must be a tuple of length 2")
# if not all(isinstance(i, int) and i >= 0 for i in coordinate):
if not all(isinstance(i, int) for i in coordinate):
raise ToolError(f"{coordinate} must be a tuple of non-negative ints")
if self.is_scaling:
x, y = self.scale_coordinates(
ScalingSource.API, coordinate[0], coordinate[1]
)
else:
x, y = coordinate
# print(f"scaled_coordinates: {x}, {y}")
# print(f"offset: {self.offset_x}, {self.offset_y}")
# x += self.offset_x # TODO - check if this is needed
# y += self.offset_y
print(f"mouse move to {x}, {y}")
if action == "mouse_move":
pyautogui.moveTo(x, y)
return ToolResult(output=f"Moved mouse to ({x}, {y})")
elif action == "left_click_drag":
current_x, current_y = pyautogui.position()
pyautogui.dragTo(x, y, duration=0.5) # Adjust duration as needed
return ToolResult(output=f"Dragged mouse from ({current_x}, {current_y}) to ({x}, {y})")
if action in ("key", "type"):
if text is None:
raise ToolError(f"text is required for {action}")
if coordinate is not None:
raise ToolError(f"coordinate is not accepted for {action}")
if not isinstance(text, str):
raise ToolError(output=f"{text} must be a string")
if action == "key":
# Handle key combinations
keys = text.split('+')
for key in keys:
key = self.key_conversion.get(key.strip(), key.strip())
key = key.lower()
pyautogui.keyDown(key) # Press down each key
for key in reversed(keys):
key = self.key_conversion.get(key.strip(), key.strip())
key = key.lower()
pyautogui.keyUp(key) # Release each key in reverse order
return ToolResult(output=f"Pressed keys: {text}")
elif action == "type":
pyautogui.typewrite(text, interval=TYPING_DELAY_MS / 1000) # Convert ms to seconds
pyautogui.press('enter')
screenshot_base64 = (await self.screenshot()).base64_image
return ToolResult(output=text, base64_image=screenshot_base64)
if action in (
"left_click",
"right_click",
"double_click",
"middle_click",
"screenshot",
"cursor_position",
"left_press",
):
if text is not None:
raise ToolError(f"text is not accepted for {action}")
if coordinate is not None:
raise ToolError(f"coordinate is not accepted for {action}")
if action == "screenshot":
return await self.screenshot()
elif action == "cursor_position":
x, y = pyautogui.position()
x, y = self.scale_coordinates(ScalingSource.COMPUTER, x, y)
return ToolResult(output=f"X={x},Y={y}")
else:
if action == "left_click":
pyautogui.click()
elif action == "right_click":
pyautogui.rightClick()
elif action == "middle_click":
pyautogui.middleClick()
elif action == "double_click":
pyautogui.doubleClick()
elif action == "left_press":
pyautogui.mouseDown()
time.sleep(1)
pyautogui.mouseUp()
return ToolResult(output=f"Performed {action}")
raise ToolError(f"Invalid action: {action}")
async def screenshot(self):
import time
time.sleep(1)
"""Take a screenshot of the current screen and return a ToolResult with the base64 encoded image."""
output_dir = Path(OUTPUT_DIR)
output_dir.mkdir(parents=True, exist_ok=True)
path = output_dir / f"screenshot_{uuid4().hex}.png"
ImageGrab.grab = partial(ImageGrab.grab, all_screens=True)
# Detect platform
system = platform.system()
if system == "Windows":
# Windows: Use screeninfo to get monitor details
screens = get_monitors()
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s.x)
if self.selected_screen < 0 or self.selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
screen = sorted_screens[self.selected_screen]
bbox = (screen.x, screen.y, screen.x + screen.width, screen.y + screen.height)
elif system == "Darwin": # macOS
# macOS: Use Quartz to get monitor details
max_displays = 32 # Maximum number of displays to handle
active_displays = Quartz.CGGetActiveDisplayList(max_displays, None, None)[1]
# Get the display bounds (resolution) for each active display
screens = []
for display_id in active_displays:
bounds = Quartz.CGDisplayBounds(display_id)
screens.append({
'id': display_id,
'x': int(bounds.origin.x),
'y': int(bounds.origin.y),
'width': int(bounds.size.width),
'height': int(bounds.size.height),
'is_primary': Quartz.CGDisplayIsMain(display_id) # Check if this is the primary display
})
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s['x'])
if self.selected_screen < 0 or self.selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
screen = sorted_screens[self.selected_screen]
bbox = (screen['x'], screen['y'], screen['x'] + screen['width'], screen['y'] + screen['height'])
else: # Linux or other OS
cmd = "xrandr | grep ' primary' | awk '{print $4}'"
try:
# output = subprocess.check_output(cmd, shell=True).decode()
# resolution = output.strip().split()[0]
# width, height = map(int, resolution.split('x'))
# bbox = (0, 0, width, height) # Assuming single primary screen for simplicity
screen = get_monitors()[0]
bbox = (screen.x, screen.y, screen.x + screen.width, screen.y + screen.height)
except subprocess.CalledProcessError:
raise RuntimeError("Failed to get screen resolution on Linux.")
# Take screenshot using the bounding box
screenshot = ImageGrab.grab(bbox=bbox)
# Set offsets (for potential future use)
self.offset_x = screen['x'] if system == "Darwin" else screen.x
self.offset_y = screen['y'] if system == "Darwin" else screen.y
print(f"target_dimension {self.target_dimension}")
if not hasattr(self, 'target_dimension'):
screenshot = self.padding_image(screenshot)
self.target_dimension = MAX_SCALING_TARGETS["WXGA"]
# Resize if target_dimensions are specified
print(f"offset is {self.offset_x}, {self.offset_y}")
print(f"target_dimension is {self.target_dimension}")
screenshot = screenshot.resize((self.target_dimension["width"], self.target_dimension["height"]))
# Save the screenshot
screenshot.save(str(path))
if path.exists():
# Return a ToolResult instance instead of a dictionary
return ToolResult(base64_image=base64.b64encode(path.read_bytes()).decode())
raise ToolError(f"Failed to take screenshot: {path} does not exist.")
def padding_image(self, screenshot):
"""Pad the screenshot to 16:10 aspect ratio, when the aspect ratio is not 16:10."""
_, height = screenshot.size
new_width = height * 16 // 10
padding_image = Image.new("RGB", (new_width, height), (255, 255, 255))
# padding to top left
padding_image.paste(screenshot, (0, 0))
return padding_image
async def shell(self, command: str, take_screenshot=True) -> ToolResult:
"""Run a shell command and return the output, error, and optionally a screenshot."""
_, stdout, stderr = await run(command)
base64_image = None
if take_screenshot:
# delay to let things settle before taking a screenshot
await asyncio.sleep(self._screenshot_delay)
base64_image = (await self.screenshot()).base64_image
return ToolResult(output=stdout, error=stderr, base64_image=base64_image)
def scale_coordinates(self, source: ScalingSource, x: int, y: int):
"""Scale coordinates to a target maximum resolution."""
if not self._scaling_enabled:
return x, y
ratio = self.width / self.height
target_dimension = None
for target_name, dimension in MAX_SCALING_TARGETS.items():
# allow some error in the aspect ratio - not ratios are exactly 16:9
if abs(dimension["width"] / dimension["height"] - ratio) < 0.02:
if dimension["width"] < self.width:
target_dimension = dimension
self.target_dimension = target_dimension
# print(f"target_dimension: {target_dimension}")
break
if target_dimension is None:
# TODO: currently we force the target to be WXGA (16:10), when it cannot find a match
target_dimension = MAX_SCALING_TARGETS["WXGA"]
self.target_dimension = MAX_SCALING_TARGETS["WXGA"]
# should be less than 1
x_scaling_factor = target_dimension["width"] / self.width
y_scaling_factor = target_dimension["height"] / self.height
if source == ScalingSource.API:
if x > self.width or y > self.height:
raise ToolError(f"Coordinates {x}, {y} are out of bounds")
# scale up
return round(x / x_scaling_factor), round(y / y_scaling_factor)
# scale down
return round(x * x_scaling_factor), round(y * y_scaling_factor)
def get_screen_size(self):
if platform.system() == "Windows":
# Use screeninfo to get primary monitor on Windows
screens = get_monitors()
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s.x)
if self.selected_screen is None:
primary_monitor = next((m for m in get_monitors() if m.is_primary), None)
return primary_monitor.width, primary_monitor.height
elif self.selected_screen < 0 or self.selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
else:
screen = sorted_screens[self.selected_screen]
return screen.width, screen.height
elif platform.system() == "Darwin":
# macOS part using Quartz to get screen information
max_displays = 32 # Maximum number of displays to handle
active_displays = Quartz.CGGetActiveDisplayList(max_displays, None, None)[1]
# Get the display bounds (resolution) for each active display
screens = []
for display_id in active_displays:
bounds = Quartz.CGDisplayBounds(display_id)
screens.append({
'id': display_id,
'x': int(bounds.origin.x),
'y': int(bounds.origin.y),
'width': int(bounds.size.width),
'height': int(bounds.size.height),
'is_primary': Quartz.CGDisplayIsMain(display_id) # Check if this is the primary display
})
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s['x'])
if self.selected_screen is None:
# Find the primary monitor
primary_monitor = next((screen for screen in screens if screen['is_primary']), None)
if primary_monitor:
return primary_monitor['width'], primary_monitor['height']
else:
raise RuntimeError("No primary monitor found.")
elif self.selected_screen < 0 or self.selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
else:
# Return the resolution of the selected screen
screen = sorted_screens[self.selected_screen]
return screen['width'], screen['height']
else: # Linux or other OS
cmd = "xrandr | grep ' primary' | awk '{print $4}'"
try:
# output = subprocess.check_output(cmd, shell=True).decode()
# resolution = output.strip().split()[0]
# width, height = map(int, resolution.split('x'))
# return width, height
screen = get_monitors()[0]
return screen.width, screen.height
except subprocess.CalledProcessError:
raise RuntimeError("Failed to get screen resolution on Linux.")
def get_mouse_position(self):
# TODO: enhance this func
from AppKit import NSEvent
from Quartz import CGEventSourceCreate, kCGEventSourceStateCombinedSessionState
loc = NSEvent.mouseLocation()
# Adjust for different coordinate system
return int(loc.x), int(self.height - loc.y)
def map_keys(self, text: str):
"""Map text to cliclick key codes if necessary."""
# For simplicity, return text as is
# Implement mapping if special keys are needed
return text

290
demo/tools/edit.py
View File

@@ -1,290 +0,0 @@
from collections import defaultdict
from pathlib import Path
from typing import Literal, get_args
from anthropic.types.beta import BetaToolTextEditor20241022Param
from .base import BaseAnthropicTool, CLIResult, ToolError, ToolResult
from .run import maybe_truncate, run
Command = Literal[
"view",
"create",
"str_replace",
"insert",
"undo_edit",
]
SNIPPET_LINES: int = 4
class EditTool(BaseAnthropicTool):
"""
An filesystem editor tool that allows the agent to view, create, and edit files.
The tool parameters are defined by Anthropic and are not editable.
"""
api_type: Literal["text_editor_20241022"] = "text_editor_20241022"
name: Literal["str_replace_editor"] = "str_replace_editor"
_file_history: dict[Path, list[str]]
def __init__(self):
self._file_history = defaultdict(list)
super().__init__()
def to_params(self) -> BetaToolTextEditor20241022Param:
return {
"name": self.name,
"type": self.api_type,
}
async def __call__(
self,
*,
command: Command,
path: str,
file_text: str | None = None,
view_range: list[int] | None = None,
old_str: str | None = None,
new_str: str | None = None,
insert_line: int | None = None,
**kwargs,
):
_path = Path(path)
self.validate_path(command, _path)
if command == "view":
return await self.view(_path, view_range)
elif command == "create":
if not file_text:
raise ToolError("Parameter `file_text` is required for command: create")
self.write_file(_path, file_text)
self._file_history[_path].append(file_text)
return ToolResult(output=f"File created successfully at: {_path}")
elif command == "str_replace":
if not old_str:
raise ToolError(
"Parameter `old_str` is required for command: str_replace"
)
return self.str_replace(_path, old_str, new_str)
elif command == "insert":
if insert_line is None:
raise ToolError(
"Parameter `insert_line` is required for command: insert"
)
if not new_str:
raise ToolError("Parameter `new_str` is required for command: insert")
return self.insert(_path, insert_line, new_str)
elif command == "undo_edit":
return self.undo_edit(_path)
raise ToolError(
f'Unrecognized command {command}. The allowed commands for the {self.name} tool are: {", ".join(get_args(Command))}'
)
def validate_path(self, command: str, path: Path):
"""
Check that the path/command combination is valid.
"""
# Check if its an absolute path
if not path.is_absolute():
suggested_path = Path("") / path
raise ToolError(
f"The path {path} is not an absolute path, it should start with `/`. Maybe you meant {suggested_path}?"
)
# Check if path exists
if not path.exists() and command != "create":
raise ToolError(
f"The path {path} does not exist. Please provide a valid path."
)
if path.exists() and command == "create":
raise ToolError(
f"File already exists at: {path}. Cannot overwrite files using command `create`."
)
# Check if the path points to a directory
if path.is_dir():
if command != "view":
raise ToolError(
f"The path {path} is a directory and only the `view` command can be used on directories"
)
async def view(self, path: Path, view_range: list[int] | None = None):
"""Implement the view command"""
if path.is_dir():
if view_range:
raise ToolError(
"The `view_range` parameter is not allowed when `path` points to a directory."
)
_, stdout, stderr = await run(
rf"find {path} -maxdepth 2 -not -path '*/\.*'"
)
if not stderr:
stdout = f"Here's the files and directories up to 2 levels deep in {path}, excluding hidden items:\n{stdout}\n"
return CLIResult(output=stdout, error=stderr)
file_content = self.read_file(path)
init_line = 1
if view_range:
if len(view_range) != 2 or not all(isinstance(i, int) for i in view_range):
raise ToolError(
"Invalid `view_range`. It should be a list of two integers."
)
file_lines = file_content.split("\n")
n_lines_file = len(file_lines)
init_line, final_line = view_range
if init_line < 1 or init_line > n_lines_file:
raise ToolError(
f"Invalid `view_range`: {view_range}. It's first element `{init_line}` should be within the range of lines of the file: {[1, n_lines_file]}"
)
if final_line > n_lines_file:
raise ToolError(
f"Invalid `view_range`: {view_range}. It's second element `{final_line}` should be smaller than the number of lines in the file: `{n_lines_file}`"
)
if final_line != -1 and final_line < init_line:
raise ToolError(
f"Invalid `view_range`: {view_range}. It's second element `{final_line}` should be larger or equal than its first `{init_line}`"
)
if final_line == -1:
file_content = "\n".join(file_lines[init_line - 1 :])
else:
file_content = "\n".join(file_lines[init_line - 1 : final_line])
return CLIResult(
output=self._make_output(file_content, str(path), init_line=init_line)
)
def str_replace(self, path: Path, old_str: str, new_str: str | None):
"""Implement the str_replace command, which replaces old_str with new_str in the file content"""
# Read the file content
file_content = self.read_file(path).expandtabs()
old_str = old_str.expandtabs()
new_str = new_str.expandtabs() if new_str is not None else ""
# Check if old_str is unique in the file
occurrences = file_content.count(old_str)
if occurrences == 0:
raise ToolError(
f"No replacement was performed, old_str `{old_str}` did not appear verbatim in {path}."
)
elif occurrences > 1:
file_content_lines = file_content.split("\n")
lines = [
idx + 1
for idx, line in enumerate(file_content_lines)
if old_str in line
]
raise ToolError(
f"No replacement was performed. Multiple occurrences of old_str `{old_str}` in lines {lines}. Please ensure it is unique"
)
# Replace old_str with new_str
new_file_content = file_content.replace(old_str, new_str)
# Write the new content to the file
self.write_file(path, new_file_content)
# Save the content to history
self._file_history[path].append(file_content)
# Create a snippet of the edited section
replacement_line = file_content.split(old_str)[0].count("\n")
start_line = max(0, replacement_line - SNIPPET_LINES)
end_line = replacement_line + SNIPPET_LINES + new_str.count("\n")
snippet = "\n".join(new_file_content.split("\n")[start_line : end_line + 1])
# Prepare the success message
success_msg = f"The file {path} has been edited. "
success_msg += self._make_output(
snippet, f"a snippet of {path}", start_line + 1
)
success_msg += "Review the changes and make sure they are as expected. Edit the file again if necessary."
return CLIResult(output=success_msg)
def insert(self, path: Path, insert_line: int, new_str: str):
"""Implement the insert command, which inserts new_str at the specified line in the file content."""
file_text = self.read_file(path).expandtabs()
new_str = new_str.expandtabs()
file_text_lines = file_text.split("\n")
n_lines_file = len(file_text_lines)
if insert_line < 0 or insert_line > n_lines_file:
raise ToolError(
f"Invalid `insert_line` parameter: {insert_line}. It should be within the range of lines of the file: {[0, n_lines_file]}"
)
new_str_lines = new_str.split("\n")
new_file_text_lines = (
file_text_lines[:insert_line]
+ new_str_lines
+ file_text_lines[insert_line:]
)
snippet_lines = (
file_text_lines[max(0, insert_line - SNIPPET_LINES) : insert_line]
+ new_str_lines
+ file_text_lines[insert_line : insert_line + SNIPPET_LINES]
)
new_file_text = "\n".join(new_file_text_lines)
snippet = "\n".join(snippet_lines)
self.write_file(path, new_file_text)
self._file_history[path].append(file_text)
success_msg = f"The file {path} has been edited. "
success_msg += self._make_output(
snippet,
"a snippet of the edited file",
max(1, insert_line - SNIPPET_LINES + 1),
)
success_msg += "Review the changes and make sure they are as expected (correct indentation, no duplicate lines, etc). Edit the file again if necessary."
return CLIResult(output=success_msg)
def undo_edit(self, path: Path):
"""Implement the undo_edit command."""
if not self._file_history[path]:
raise ToolError(f"No edit history found for {path}.")
old_text = self._file_history[path].pop()
self.write_file(path, old_text)
return CLIResult(
output=f"Last edit to {path} undone successfully. {self._make_output(old_text, str(path))}"
)
def read_file(self, path: Path):
"""Read the content of a file from a given path; raise a ToolError if an error occurs."""
try:
return path.read_text()
except Exception as e:
raise ToolError(f"Ran into {e} while trying to read {path}") from None
def write_file(self, path: Path, file: str):
"""Write the content of a file to a given path; raise a ToolError if an error occurs."""
try:
path.write_text(file)
except Exception as e:
raise ToolError(f"Ran into {e} while trying to write to {path}") from None
def _make_output(
self,
file_content: str,
file_descriptor: str,
init_line: int = 1,
expand_tabs: bool = True,
):
"""Generate output for the CLI based on the content of a file."""
file_content = maybe_truncate(file_content)
if expand_tabs:
file_content = file_content.expandtabs()
file_content = "\n".join(
[
f"{i + init_line:6}\t{line}"
for i, line in enumerate(file_content.split("\n"))
]
)
return (
f"Here's the result of running `cat -n` on {file_descriptor}:\n"
+ file_content
+ "\n"
)

42
demo/tools/run.py
View File

@@ -1,42 +0,0 @@
"""Utility to run shell commands asynchronously with a timeout."""
import asyncio
TRUNCATED_MESSAGE: str = "<response clipped><NOTE>To save on context only part of this file has been shown to you. You should retry this tool after you have searched inside the file with `grep -n` in order to find the line numbers of what you are looking for.</NOTE>"
MAX_RESPONSE_LEN: int = 16000
def maybe_truncate(content: str, truncate_after: int | None = MAX_RESPONSE_LEN):
"""Truncate content and append a notice if content exceeds the specified length."""
return (
content
if not truncate_after or len(content) <= truncate_after
else content[:truncate_after] + TRUNCATED_MESSAGE
)
async def run(
cmd: str,
timeout: float | None = 120.0, # seconds
truncate_after: int | None = MAX_RESPONSE_LEN,
):
"""Run a shell command asynchronously with a timeout."""
process = await asyncio.create_subprocess_shell(
cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE
)
try:
stdout, stderr = await asyncio.wait_for(process.communicate(), timeout=timeout)
return (
process.returncode or 0,
maybe_truncate(stdout.decode(), truncate_after=truncate_after),
maybe_truncate(stderr.decode(), truncate_after=truncate_after),
)
except asyncio.TimeoutError as exc:
try:
process.kill()
except ProcessLookupError:
pass
raise TimeoutError(
f"Command '{cmd}' timed out after {timeout} seconds"
) from exc

185
demo/tools/screen_capture.py
View File

@@ -1,185 +0,0 @@
import subprocess
import base64
from pathlib import Path
from PIL import ImageGrab
from uuid import uuid4
from screeninfo import get_monitors
import platform
if platform.system() == "Darwin":
import Quartz # uncomment this line if you are on macOS
from PIL import ImageGrab
from functools import partial
from .base import BaseAnthropicTool, ToolError, ToolResult
OUTPUT_DIR = "./tmp/outputs"
def get_screenshot(selected_screen: int = 0, resize: bool = True, target_width: int = 1920, target_height: int = 1080):
# print(f"get_screenshot selected_screen: {selected_screen}")
# Get screen width and height using Windows command
display_num = None
offset_x = 0
offset_y = 0
selected_screen = selected_screen
width, height = _get_screen_size()
"""Take a screenshot of the current screen and return a ToolResult with the base64 encoded image."""
output_dir = Path(OUTPUT_DIR)
output_dir.mkdir(parents=True, exist_ok=True)
path = output_dir / f"screenshot_{uuid4().hex}.png"
ImageGrab.grab = partial(ImageGrab.grab, all_screens=True)
# Detect platform
system = platform.system()
if system == "Windows":
# Windows: Use screeninfo to get monitor details
screens = get_monitors()
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s.x)
if selected_screen < 0 or selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
screen = sorted_screens[selected_screen]
bbox = (screen.x, screen.y, screen.x + screen.width, screen.y + screen.height)
elif system == "Darwin": # macOS
# macOS: Use Quartz to get monitor details
max_displays = 32 # Maximum number of displays to handle
active_displays = Quartz.CGGetActiveDisplayList(max_displays, None, None)[1]
# Get the display bounds (resolution) for each active display
screens = []
for display_id in active_displays:
bounds = Quartz.CGDisplayBounds(display_id)
screens.append({
'id': display_id,
'x': int(bounds.origin.x),
'y': int(bounds.origin.y),
'width': int(bounds.size.width),
'height': int(bounds.size.height),
'is_primary': Quartz.CGDisplayIsMain(display_id) # Check if this is the primary display
})
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s['x'])
# print(f"Darwin sorted_screens: {sorted_screens}")
if selected_screen < 0 or selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
screen = sorted_screens[selected_screen]
bbox = (screen['x'], screen['y'], screen['x'] + screen['width'], screen['y'] + screen['height'])
else: # Linux or other OS
cmd = "xrandr | grep ' primary' | awk '{print $4}'"
try:
# output = subprocess.check_output(cmd, shell=True).decode()
# resolution = output.strip().split()[0]
# width, height = map(int, resolution.split('x'))
screen = get_monitors()[0]
width, height = screen.width, screen.height
bbox = (0, 0, width, height) # Assuming single primary screen for simplicity
except subprocess.CalledProcessError:
raise RuntimeError("Failed to get screen resolution on Linux.")
# Take screenshot using the bounding box
screenshot = ImageGrab.grab(bbox=bbox)
import os
if (display_num := os.getenv("DISPLAY_NUM")) is not None:
display_num = int(display_num)
_display_prefix = f"DISPLAY=:{display_num} "
else:
display_num = None
_display_prefix = ""
screenshot_cmd = f"{_display_prefix}scrot -p {path}"
import pdb; pdb.set_trace()
result = subprocess.run(screenshot_cmd, shell=True, capture_output=True)
# Set offsets (for potential future use)
offset_x = screen['x'] if system == "Darwin" else screen.x
offset_y = screen['y'] if system == "Darwin" else screen.y
# # Resize if
if resize:
screenshot = screenshot.resize((target_width, target_height))
# Save the screenshot
# screenshot.save(str(path))
if path.exists():
# Return a ToolResult instance instead of a dictionary
return screenshot, path
raise ToolError(f"Failed to take screenshot: {path} does not exist.")
def _get_screen_size(selected_screen: int = 0):
if platform.system() == "Windows":
# Use screeninfo to get primary monitor on Windows
screens = get_monitors()
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s.x)
if selected_screen is None:
primary_monitor = next((m for m in get_monitors() if m.is_primary), None)
return primary_monitor.width, primary_monitor.height
elif selected_screen < 0 or selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
else:
screen = sorted_screens[selected_screen]
return screen.width, screen.height
elif platform.system() == "Darwin":
# macOS part using Quartz to get screen information
max_displays = 32 # Maximum number of displays to handle
active_displays = Quartz.CGGetActiveDisplayList(max_displays, None, None)[1]
# Get the display bounds (resolution) for each active display
screens = []
for display_id in active_displays:
bounds = Quartz.CGDisplayBounds(display_id)
screens.append({
'id': display_id,
'x': int(bounds.origin.x),
'y': int(bounds.origin.y),
'width': int(bounds.size.width),
'height': int(bounds.size.height),
'is_primary': Quartz.CGDisplayIsMain(display_id) # Check if this is the primary display
})
# Sort screens by x position to arrange from left to right
sorted_screens = sorted(screens, key=lambda s: s['x'])
if selected_screen is None:
# Find the primary monitor
primary_monitor = next((screen for screen in screens if screen['is_primary']), None)
if primary_monitor:
return primary_monitor['width'], primary_monitor['height']
else:
raise RuntimeError("No primary monitor found.")
elif selected_screen < 0 or selected_screen >= len(screens):
raise IndexError("Invalid screen index.")
else:
# Return the resolution of the selected screen
screen = sorted_screens[selected_screen]
return screen['width'], screen['height']
else: # Linux or other OS
cmd = "xrandr | grep ' primary' | awk '{print $4}'"
try:
# output = subprocess.check_output(cmd, shell=True).decode()
# resolution = output.strip().split()[0]
# width, height = map(int, resolution.split('x'))
# return width, height
screen = get_monitors()[0]
return screen.width, screen.height
except subprocess.CalledProcessError:
raise RuntimeError("Failed to get screen resolution on Linux.")

6
gradio_demo.py
View File

@@ -8,7 +8,7 @@ import io
import base64, os
from utils import check_ocr_box, get_yolo_model, get_caption_model_processor, get_som_labeled_img
from util.utils import check_ocr_box, get_yolo_model, get_caption_model_processor, get_som_labeled_img
import torch
from PIL import Image
@@ -17,8 +17,6 @@ yolo_model = get_yolo_model(model_path='weights/icon_detect_v1_5/best.pt')
caption_model_processor = get_caption_model_processor(model_name="florence2", model_name_or_path="weights/icon_caption_florence")
# caption_model_processor = get_caption_model_processor(model_name="blip2", model_name_or_path="weights/icon_caption_blip2")
MARKDOWN = """
# OmniParser for Pure Vision Based General GUI Agent 🔥
<div>
@@ -65,8 +63,6 @@ def process(
# parsed_content_list = str(parsed_content_list)
return image, str(parsed_content_list)
with gr.Blocks() as demo:
gr.Markdown(MARKDOWN)
with gr.Row():

BIN
imgs/gradioicon.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 33 KiB

BIN
imgs/header_bar.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 210 KiB

BIN
imgs/header_bar_thin.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 132 KiB

BIN
imgs/omniboxicon.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 3.4 KiB

BIN
imgs/omniparsericon.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 7.4 KiB

BIN
imgs/som_overlaid_omni.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 279 KiB

60
omniparser.py
View File

@@ -1,60 +0,0 @@
from utils import get_som_labeled_img, check_ocr_box, get_yolo_model
import torch
from ultralytics import YOLO
from PIL import Image
from typing import Dict, Tuple, List
import io
import base64
config = {
'som_model_path': 'finetuned_icon_detect.pt',
'device': 'cpu',
'caption_model_path': 'Salesforce/blip2-opt-2.7b',
'draw_bbox_config': {
'text_scale': 0.8,
'text_thickness': 2,
'text_padding': 3,
'thickness': 3,
},
'BOX_TRESHOLD': 0.05
}
class Omniparser(object):
def __init__(self, config: Dict):
self.config = config
self.som_model = get_yolo_model(model_path=config['som_model_path'])
# self.caption_model_processor = get_caption_model_processor(config['caption_model_path'], device=cofig['device'])
# self.caption_model_processor['model'].to(torch.float32)
def parse(self, image_path: str):
print('Parsing image:', image_path)
ocr_bbox_rslt, is_goal_filtered = check_ocr_box(image_path, display_img = False, output_bb_format='xyxy', goal_filtering=None, easyocr_args={'paragraph': False, 'text_threshold':0.9})
text, ocr_bbox = ocr_bbox_rslt
draw_bbox_config = self.config['draw_bbox_config']
BOX_TRESHOLD = self.config['BOX_TRESHOLD']
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(image_path, self.som_model, BOX_TRESHOLD = BOX_TRESHOLD, output_coord_in_ratio=False, ocr_bbox=ocr_bbox,draw_bbox_config=draw_bbox_config, caption_model_processor=None, ocr_text=text,use_local_semantics=False)
image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
# formating output
return_list = [{'from': 'omniparser', 'shape': {'x':coord[0], 'y':coord[1], 'width':coord[2], 'height':coord[3]},
'text': parsed_content_list[i].split(': ')[1], 'type':'text'} for i, (k, coord) in enumerate(label_coordinates.items()) if i < len(parsed_content_list)]
return_list.extend(
[{'from': 'omniparser', 'shape': {'x':coord[0], 'y':coord[1], 'width':coord[2], 'height':coord[3]},
'text': 'None', 'type':'icon'} for i, (k, coord) in enumerate(label_coordinates.items()) if i >= len(parsed_content_list)]
)
return [image, return_list]
parser = Omniparser(config)
image_path = 'examples/pc_1.png'
# time the parser
import time
s = time.time()
image, parsed_content_list = parser.parse(image_path)
device = config['device']
print(f'Time taken for Omniparser on {device}:', time.time() - s)

1
omnitool/gradio/.gitignore vendored Normal file
View File

@@ -0,0 +1 @@
tmp/

0
demo/__init__.py → omnitool/gradio/__init__.py
View File

369
demo/gui_agent/anthropic_agent.py → omnitool/gradio/agent/anthropic_agent.py
View File

@@ -1,207 +1,162 @@
"""
Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
"""
import asyncio
import platform
from collections.abc import Callable
from datetime import datetime
from enum import StrEnum
from typing import Any, cast
from anthropic import Anthropic, AnthropicBedrock, AnthropicVertex, APIResponse
from anthropic.types import (
ToolResultBlockParam,
)
from anthropic.types.beta import (
BetaContentBlock,
BetaContentBlockParam,
BetaImageBlockParam,
BetaMessage,
BetaMessageParam,
BetaTextBlockParam,
BetaToolResultBlockParam,
)
from anthropic.types import TextBlock
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
from tools import BashTool, ComputerTool, EditTool, ToolCollection, ToolResult
from PIL import Image
from io import BytesIO
import gradio as gr
from typing import Dict
BETA_FLAG = "computer-use-2024-10-22"
class APIProvider(StrEnum):
ANTHROPIC = "anthropic"
BEDROCK = "bedrock"
VERTEX = "vertex"
PROVIDER_TO_DEFAULT_MODEL_NAME: dict[APIProvider, str] = {
APIProvider.ANTHROPIC: "claude-3-5-sonnet-20241022",
APIProvider.BEDROCK: "anthropic.claude-3-5-sonnet-20241022-v2:0",
APIProvider.VERTEX: "claude-3-5-sonnet-v2@20241022",
}
# Check OS
SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
* You are utilizing a Windows system with internet access.
* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
</SYSTEM_CAPABILITY>
"""
class AnthropicActor:
def __init__(
self,
model: str,
provider: APIProvider,
system_prompt_suffix: str,
api_key: str,
api_response_callback: Callable[[APIResponse[BetaMessage]], None],
max_tokens: int = 4096,
only_n_most_recent_images: int | None = None,
selected_screen: int = 0,
print_usage: bool = True,
):
self.model = model
self.provider = provider
self.system_prompt_suffix = system_prompt_suffix
self.api_key = api_key
self.api_response_callback = api_response_callback
self.max_tokens = max_tokens
self.only_n_most_recent_images = only_n_most_recent_images
self.selected_screen = selected_screen
self.tool_collection = ToolCollection(
ComputerTool(selected_screen=selected_screen),
BashTool(),
EditTool(),
)
self.system = (
f"{SYSTEM_PROMPT}{' ' + system_prompt_suffix if system_prompt_suffix else ''}"
)
self.total_token_usage = 0
self.total_cost = 0
self.print_usage = print_usage
# Instantiate the appropriate API client based on the provider
if provider == APIProvider.ANTHROPIC:
self.client = Anthropic(api_key=api_key)
elif provider == APIProvider.VERTEX:
self.client = AnthropicVertex()
elif provider == APIProvider.BEDROCK:
self.client = AnthropicBedrock()
def __call__(
self,
*,
messages: list[BetaMessageParam]
):
"""
Generate a response given history messages.
"""
if self.only_n_most_recent_images:
_maybe_filter_to_n_most_recent_images(messages, self.only_n_most_recent_images)
# Call the API synchronously
raw_response = self.client.beta.messages.with_raw_response.create(
max_tokens=self.max_tokens,
messages=messages,
model=self.model,
system=self.system,
tools=self.tool_collection.to_params(),
betas=["computer-use-2024-10-22"],
)
self.api_response_callback(cast(APIResponse[BetaMessage], raw_response))
response = raw_response.parse()
print(f"AnthropicActor response: {response}")
self.total_token_usage += response.usage.input_tokens + response.usage.output_tokens
self.total_cost += (response.usage.input_tokens * 3 / 1000000 + response.usage.output_tokens * 15 / 1000000)
if self.print_usage:
print(f"Claude total token usage so far: {self.total_token_usage}, total cost so far: $USD{self.total_cost}")
return response
def _maybe_filter_to_n_most_recent_images(
messages: list[BetaMessageParam],
images_to_keep: int,
min_removal_threshold: int = 10,
):
"""
With the assumption that images are screenshots that are of diminishing value as
the conversation progresses, remove all but the final `images_to_keep` tool_result
images in place, with a chunk of min_removal_threshold to reduce the amount we
break the implicit prompt cache.
"""
if images_to_keep is None:
return messages
tool_result_blocks = cast(
list[ToolResultBlockParam],
[
item
for message in messages
for item in (
message["content"] if isinstance(message["content"], list) else []
)
if isinstance(item, dict) and item.get("type") == "tool_result"
],
)
total_images = sum(
1
for tool_result in tool_result_blocks
for content in tool_result.get("content", [])
if isinstance(content, dict) and content.get("type") == "image"
)
images_to_remove = total_images - images_to_keep
# for better cache behavior, we want to remove in chunks
images_to_remove -= images_to_remove % min_removal_threshold
for tool_result in tool_result_blocks:
if isinstance(tool_result.get("content"), list):
new_content = []
for content in tool_result.get("content", []):
if isinstance(content, dict) and content.get("type") == "image":
if images_to_remove > 0:
images_to_remove -= 1
continue
new_content.append(content)
tool_result["content"] = new_content
if __name__ == "__main__":
pass
# client = Anthropic(api_key="")
# response = client.beta.messages.with_raw_response.create(
# max_tokens=4096,
# model="claude-3-5-sonnet-20241022",
# system=SYSTEM_PROMPT,
# # tools=ToolCollection(
# # ComputerTool(selected_screen=0),
# # BashTool(),
# # EditTool(),
# # ).to_params(),
# betas=["computer-use-2024-10-22"],
# messages=[
# {"role": "user", "content": "click on (199, 199)."}
# ],
# )
# print(f"AnthropicActor response: {response.parse().usage.input_tokens+response.parse().usage.output_tokens}")
"""
Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
"""
import asyncio
import platform
from collections.abc import Callable
from datetime import datetime
from enum import StrEnum
from typing import Any, cast
from anthropic import Anthropic, AnthropicBedrock, AnthropicVertex, APIResponse
from anthropic.types import (
ToolResultBlockParam,
)
from anthropic.types.beta import (
BetaContentBlock,
BetaContentBlockParam,
BetaImageBlockParam,
BetaMessage,
BetaMessageParam,
BetaTextBlockParam,
BetaToolResultBlockParam,
)
from anthropic.types import TextBlock
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
from tools import ComputerTool, ToolCollection, ToolResult
from PIL import Image
from io import BytesIO
import gradio as gr
from typing import Dict
BETA_FLAG = "computer-use-2024-10-22"
class APIProvider(StrEnum):
ANTHROPIC = "anthropic"
BEDROCK = "bedrock"
VERTEX = "vertex"
SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
* You are utilizing a Windows system with internet access.
* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
</SYSTEM_CAPABILITY>
"""
class AnthropicActor:
def __init__(
self,
model: str,
provider: APIProvider,
api_key: str,
api_response_callback: Callable[[APIResponse[BetaMessage]], None],
max_tokens: int = 4096,
only_n_most_recent_images: int | None = None,
print_usage: bool = True,
):
self.model = model
self.provider = provider
self.api_key = api_key
self.api_response_callback = api_response_callback
self.max_tokens = max_tokens
self.only_n_most_recent_images = only_n_most_recent_images
self.tool_collection = ToolCollection(ComputerTool())
self.system = SYSTEM_PROMPT
self.total_token_usage = 0
self.total_cost = 0
self.print_usage = print_usage
# Instantiate the appropriate API client based on the provider
if provider == APIProvider.ANTHROPIC:
self.client = Anthropic(api_key=api_key)
elif provider == APIProvider.VERTEX:
self.client = AnthropicVertex()
elif provider == APIProvider.BEDROCK:
self.client = AnthropicBedrock()
def __call__(
self,
*,
messages: list[BetaMessageParam]
):
"""
Generate a response given history messages.
"""
if self.only_n_most_recent_images:
_maybe_filter_to_n_most_recent_images(messages, self.only_n_most_recent_images)
# Call the API synchronously
raw_response = self.client.beta.messages.with_raw_response.create(
max_tokens=self.max_tokens,
messages=messages,
model=self.model,
system=self.system,
tools=self.tool_collection.to_params(),
betas=["computer-use-2024-10-22"],
)
self.api_response_callback(cast(APIResponse[BetaMessage], raw_response))
response = raw_response.parse()
print(f"AnthropicActor response: {response}")
self.total_token_usage += response.usage.input_tokens + response.usage.output_tokens
self.total_cost += (response.usage.input_tokens * 3 / 1000000 + response.usage.output_tokens * 15 / 1000000)
if self.print_usage:
print(f"Claude total token usage so far: {self.total_token_usage}, total cost so far: $USD{self.total_cost}")
return response
def _maybe_filter_to_n_most_recent_images(
messages: list[BetaMessageParam],
images_to_keep: int,
min_removal_threshold: int = 10,
):
"""
With the assumption that images are screenshots that are of diminishing value as
the conversation progresses, remove all but the final `images_to_keep` tool_result
images in place, with a chunk of min_removal_threshold to reduce the amount we
break the implicit prompt cache.
"""
if images_to_keep is None:
return messages
tool_result_blocks = cast(
list[ToolResultBlockParam],
[
item
for message in messages
for item in (
message["content"] if isinstance(message["content"], list) else []
)
if isinstance(item, dict) and item.get("type") == "tool_result"
],
)
total_images = sum(
1
for tool_result in tool_result_blocks
for content in tool_result.get("content", [])
if isinstance(content, dict) and content.get("type") == "image"
)
images_to_remove = total_images - images_to_keep
# for better cache behavior, we want to remove in chunks
images_to_remove -= images_to_remove % min_removal_threshold
for tool_result in tool_result_blocks:
if isinstance(tool_result.get("content"), list):
new_content = []
for content in tool_result.get("content", []):
if isinstance(content, dict) and content.get("type") == "image":
if images_to_remove > 0:
images_to_remove -= 1
continue
new_content.append(content)
tool_result["content"] = new_content

59
omnitool/gradio/agent/llm_utils/groqclient.py Normal file
View File

@@ -0,0 +1,59 @@
from groq import Groq
import os
from .utils import is_image_path
def run_groq_interleaved(messages: list, system: str, model_name: str, api_key: str, max_tokens=256, temperature=0.6):
"""
Run a chat completion through Groq's API, ignoring any images in the messages.
"""
api_key = api_key or os.environ.get("GROQ_API_KEY")
if not api_key:
raise ValueError("GROQ_API_KEY is not set")
client = Groq(api_key=api_key)
# avoid using system messages for R1
final_messages = [{"role": "user", "content": system}]
if isinstance(messages, list):
for item in messages:
if isinstance(item, dict):
# For dict items, concatenate all text content, ignoring images
text_contents = []
for cnt in item["content"]:
if isinstance(cnt, str):
if not is_image_path(cnt): # Skip image paths
text_contents.append(cnt)
else:
text_contents.append(str(cnt))
if text_contents: # Only add if there's text content
message = {"role": "user", "content": " ".join(text_contents)}
final_messages.append(message)
else: # str
message = {"role": "user", "content": item}
final_messages.append(message)
elif isinstance(messages, str):
final_messages.append({"role": "user", "content": messages})
try:
completion = client.chat.completions.create(
model="deepseek-r1-distill-llama-70b",
messages=final_messages,
temperature=0.6,
max_completion_tokens=max_tokens,
top_p=0.95,
stream=False,
reasoning_format="raw"
)
response = completion.choices[0].message.content
final_answer = response.split('</think>\n')[-1] if '</think>' in response else response
final_answer = final_answer.replace("<output>", "").replace("</output>", "")
token_usage = completion.usage.total_tokens
return final_answer, token_usage
except Exception as e:
print(f"Error in interleaved Groq: {e}")
return str(e), 0

62
omnitool/gradio/agent/llm_utils/oaiclient.py Normal file
View File

@@ -0,0 +1,62 @@
import os
import logging
import base64
import requests
from .utils import is_image_path, encode_image
def run_oai_interleaved(messages: list, system: str, model_name: str, api_key: str, max_tokens=256, temperature=0, provider_base_url: str = "https://api.openai.com/v1"):
headers = {"Content-Type": "application/json",
"Authorization": f"Bearer {api_key}"}
final_messages = [{"role": "system", "content": system}]
if type(messages) == list:
for item in messages:
contents = []
if isinstance(item, dict):
for cnt in item["content"]:
if isinstance(cnt, str):
if is_image_path(cnt) and 'o3-mini' not in model_name:
# 03 mini does not support images
base64_image = encode_image(cnt)
content = {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}
else:
content = {"type": "text", "text": cnt}
else:
# in this case it is a text block from anthropic
content = {"type": "text", "text": str(cnt)}
contents.append(content)
message = {"role": 'user', "content": contents}
else: # str
contents.append({"type": "text", "text": item})
message = {"role": "user", "content": contents}
final_messages.append(message)
elif isinstance(messages, str):
final_messages = [{"role": "user", "content": messages}]
payload = {
"model": model_name,
"messages": final_messages,
}
if 'o1' in model_name or 'o3-mini' in model_name:
payload['reasoning_effort'] = 'low'
payload['max_completion_tokens'] = max_tokens
else:
payload['max_tokens'] = max_tokens
response = requests.post(
f"{provider_base_url}/chat/completions", headers=headers, json=payload
)
try:
text = response.json()['choices'][0]['message']['content']
token_usage = int(response.json()['usage']['total_tokens'])
return text, token_usage
except Exception as e:
print(f"Error in interleaved openAI: {e}. This may due to your invalid API key. Please check the response: {response.json()} ")
return response.json()

44
omnitool/gradio/agent/llm_utils/omniparserclient.py Normal file
View File

@@ -0,0 +1,44 @@
import requests
import base64
from pathlib import Path
from tools.screen_capture import get_screenshot
from agent.llm_utils.utils import encode_image
OUTPUT_DIR = "./tmp/outputs"
class OmniParserClient:
def __init__(self,
url: str) -> None:
self.url = url
def __call__(self,):
screenshot, screenshot_path = get_screenshot()
screenshot_path = str(screenshot_path)
image_base64 = encode_image(screenshot_path)
response = requests.post(self.url, json={"base64_image": image_base64})
response_json = response.json()
print('omniparser latency:', response_json['latency'])
som_image_data = base64.b64decode(response_json['som_image_base64'])
screenshot_path_uuid = Path(screenshot_path).stem.replace("screenshot_", "")
som_screenshot_path = f"{OUTPUT_DIR}/screenshot_som_{screenshot_path_uuid}.png"
with open(som_screenshot_path, "wb") as f:
f.write(som_image_data)
response_json['width'] = screenshot.size[0]
response_json['height'] = screenshot.size[1]
response_json['original_screenshot_base64'] = image_base64
response_json['screenshot_uuid'] = screenshot_path_uuid
response_json = self.reformat_messages(response_json)
return response_json
def reformat_messages(self, response_json: dict):
screen_info = ""
for idx, element in enumerate(response_json["parsed_content_list"]):
element['idx'] = idx
if element['type'] == 'text':
screen_info += f'ID: {idx}, Text: {element["content"]}\n'
elif element['type'] == 'icon':
screen_info += f'ID: {idx}, Icon: {element["content"]}\n'
response_json['screen_info'] = screen_info
return response_json

13
omnitool/gradio/agent/llm_utils/utils.py Normal file
View File

@@ -0,0 +1,13 @@
import base64
def is_image_path(text):
image_extensions = (".jpg", ".jpeg", ".png", ".gif", ".bmp", ".tiff", ".tif")
if text.endswith(image_extensions):
return True
else:
return False
def encode_image(image_path):
"""Encode image file to base64."""
with open(image_path, "rb") as image_file:
return base64.b64encode(image_file.read()).decode("utf-8")

338
omnitool/gradio/agent/vlm_agent.py Normal file
View File

@@ -0,0 +1,338 @@
import json
from collections.abc import Callable
from typing import cast, Callable
import uuid
from PIL import Image, ImageDraw
import base64
from io import BytesIO
from anthropic import APIResponse
from anthropic.types import ToolResultBlockParam
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock, BetaMessageParam, BetaUsage
from agent.llm_utils.oaiclient import run_oai_interleaved
from agent.llm_utils.groqclient import run_groq_interleaved
from agent.llm_utils.utils import is_image_path
import time
import re
OUTPUT_DIR = "./tmp/outputs"
def extract_data(input_string, data_type):
# Regular expression to extract content starting from '```python' until the end if there are no closing backticks
pattern = f"```{data_type}" + r"(.*?)(```|$)"
# Extract content
# re.DOTALL allows '.' to match newlines as well
matches = re.findall(pattern, input_string, re.DOTALL)
# Return the first match if exists, trimming whitespace and ignoring potential closing backticks
return matches[0][0].strip() if matches else input_string
class VLMAgent:
def __init__(
self,
model: str,
provider: str,
api_key: str,
output_callback: Callable,
api_response_callback: Callable,
max_tokens: int = 4096,
only_n_most_recent_images: int | None = None,
print_usage: bool = True,
):
if model == "omniparser + gpt-4o":
self.model = "gpt-4o-2024-11-20"
elif model == "omniparser + R1":
self.model = "deepseek-r1-distill-llama-70b"
elif model == "omniparser + qwen2.5vl":
self.model = "qwen2.5-vl-72b-instruct"
elif model == "omniparser + o1":
self.model = "o1"
elif model == "omniparser + o3-mini":
self.model = "o3-mini"
else:
raise ValueError(f"Model {model} not supported")
self.provider = provider
self.api_key = api_key
self.api_response_callback = api_response_callback
self.max_tokens = max_tokens
self.only_n_most_recent_images = only_n_most_recent_images
self.output_callback = output_callback
self.print_usage = print_usage
self.total_token_usage = 0
self.total_cost = 0
self.step_count = 0
self.system = ''
def __call__(self, messages: list, parsed_screen: list[str, list, dict]):
self.step_count += 1
image_base64 = parsed_screen['original_screenshot_base64']
latency_omniparser = parsed_screen['latency']
self.output_callback(f'-- Step {self.step_count}: --', sender="bot")
screen_info = str(parsed_screen['screen_info'])
screenshot_uuid = parsed_screen['screenshot_uuid']
screen_width, screen_height = parsed_screen['width'], parsed_screen['height']
boxids_and_labels = parsed_screen["screen_info"]
system = self._get_system_prompt(boxids_and_labels)
# drop looping actions msg, byte image etc
planner_messages = messages
_remove_som_images(planner_messages)
_maybe_filter_to_n_most_recent_images(planner_messages, self.only_n_most_recent_images)
if isinstance(planner_messages[-1], dict):
if not isinstance(planner_messages[-1]["content"], list):
planner_messages[-1]["content"] = [planner_messages[-1]["content"]]
planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_{screenshot_uuid}.png")
planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_som_{screenshot_uuid}.png")
start = time.time()
if "gpt" in self.model or "o1" in self.model or "o3-mini" in self.model:
vlm_response, token_usage = run_oai_interleaved(
messages=planner_messages,
system=system,
model_name=self.model,
api_key=self.api_key,
max_tokens=self.max_tokens,
provider_base_url="https://api.openai.com/v1",
temperature=0,
)
print(f"oai token usage: {token_usage}")
self.total_token_usage += token_usage
if 'gpt' in self.model:
self.total_cost += (token_usage * 2.5 / 1000000) # https://openai.com/api/pricing/
elif 'o1' in self.model:
self.total_cost += (token_usage * 15 / 1000000) # https://openai.com/api/pricing/
elif 'o3-mini' in self.model:
self.total_cost += (token_usage * 1.1 / 1000000) # https://openai.com/api/pricing/
elif "r1" in self.model:
vlm_response, token_usage = run_groq_interleaved(
messages=planner_messages,
system=system,
model_name=self.model,
api_key=self.api_key,
max_tokens=self.max_tokens,
)
print(f"groq token usage: {token_usage}")
self.total_token_usage += token_usage
self.total_cost += (token_usage * 0.99 / 1000000)
elif "qwen" in self.model:
vlm_response, token_usage = run_oai_interleaved(
messages=planner_messages,
system=system,
model_name=self.model,
api_key=self.api_key,
max_tokens=min(2048, self.max_tokens),
provider_base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
temperature=0,
)
print(f"qwen token usage: {token_usage}")
self.total_token_usage += token_usage
self.total_cost += (token_usage * 2.2 / 1000000) # https://help.aliyun.com/zh/model-studio/getting-started/models?spm=a2c4g.11186623.0.0.74b04823CGnPv7#fe96cfb1a422a
else:
raise ValueError(f"Model {self.model} not supported")
latency_vlm = time.time() - start
self.output_callback(f"LLM: {latency_vlm:.2f}s, OmniParser: {latency_omniparser:.2f}s", sender="bot")
print(f"{vlm_response}")
if self.print_usage:
print(f"Total token so far: {self.total_token_usage}. Total cost so far: $USD{self.total_cost:.5f}")
vlm_response_json = extract_data(vlm_response, "json")
vlm_response_json = json.loads(vlm_response_json)
img_to_show_base64 = parsed_screen["som_image_base64"]
if "Box ID" in vlm_response_json:
bbox = parsed_screen["parsed_content_list"][int(vlm_response_json["Box ID"])]["bbox"]
vlm_response_json["box_centroid_coordinate"] = [int((bbox[0] + bbox[2]) / 2 * screen_width), int((bbox[1] + bbox[3]) / 2 * screen_height)]
img_to_show_data = base64.b64decode(img_to_show_base64)
img_to_show = Image.open(BytesIO(img_to_show_data))
draw = ImageDraw.Draw(img_to_show)
x, y = vlm_response_json["box_centroid_coordinate"]
radius = 10
draw.ellipse((x - radius, y - radius, x + radius, y + radius), fill='red')
draw.ellipse((x - radius*3, y - radius*3, x + radius*3, y + radius*3), fill=None, outline='red', width=2)
buffered = BytesIO()
img_to_show.save(buffered, format="PNG")
img_to_show_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
self.output_callback(f'<img src="data:image/png;base64,{img_to_show_base64}">', sender="bot")
self.output_callback(
f'<details>'
f' <summary>Parsed Screen elemetns by OmniParser</summary>'
f' <pre>{screen_info}</pre>'
f'</details>',
sender="bot"
)
vlm_plan_str = ""
for key, value in vlm_response_json.items():
if key == "Reasoning":
vlm_plan_str += f'{value}'
else:
vlm_plan_str += f'\n{key}: {value}'
# construct the response so that anthropicExcutor can execute the tool
response_content = [BetaTextBlock(text=vlm_plan_str, type='text')]
if 'box_centroid_coordinate' in vlm_response_json:
move_cursor_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
input={'action': 'mouse_move', 'coordinate': vlm_response_json["box_centroid_coordinate"]},
name='computer', type='tool_use')
response_content.append(move_cursor_block)
if vlm_response_json["Next Action"] == "None":
print("Task paused/completed.")
elif vlm_response_json["Next Action"] == "type":
sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
input={'action': vlm_response_json["Next Action"], 'text': vlm_response_json["value"]},
name='computer', type='tool_use')
response_content.append(sim_content_block)
else:
sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
input={'action': vlm_response_json["Next Action"]},
name='computer', type='tool_use')
response_content.append(sim_content_block)
response_message = BetaMessage(id=f'toolu_{uuid.uuid4()}', content=response_content, model='', role='assistant', type='message', stop_reason='tool_use', usage=BetaUsage(input_tokens=0, output_tokens=0))
return response_message, vlm_response_json
def _api_response_callback(self, response: APIResponse):
self.api_response_callback(response)
def _get_system_prompt(self, screen_info: str = ""):
main_section = f"""
You are using a Windows device.
You are able to use a mouse and keyboard to interact with the computer based on the given task and screenshot.
You can only interact with the desktop GUI (no terminal or application menu access).
You may be given some history plan and actions, this is the response from the previous loop.
You should carefully consider your plan base on the task, screenshot, and history actions.
Here is the list of all detected bounding boxes by IDs on the screen and their description:{screen_info}
Your available "Next Action" only include:
- type: types a string of text.
- left_click: move mouse to box id and left clicks.
- right_click: move mouse to box id and right clicks.
- double_click: move mouse to box id and double clicks.
- hover: move mouse to box id.
- scroll_up: scrolls the screen up.
- scroll_down: scrolls the screen down.
- wait: waits for 1 second for the device to load or respond.
Based on the visual information from the screenshot image and the detected bounding boxes, please determine the next action, the Box ID you should operate on (if action is not 'type', 'hover', 'scroll_up', 'scroll_down', 'wait'), and the value (if the action is 'type') in order to complete the task.
Output format:
```json
{{
"Reasoning": str, # describe what is in the current screen, taking into account the history, then describe your step-by-step thoughts on how to achieve the task, choose one action from available actions at a time.
"Next Action": "action_type, action description" | "None" # one action at a time, describe it in short and precisely.
"Box ID": n,
"value": "xxx" # only provide value field if the action is type, else don't include value key
}}
```
One Example:
```json
{{
"Reasoning": "The current screen shows google result of amazon, in previous action I have searched amazon on google. Then I need to click on the first search results to go to amazon.com.",
"Next Action": "left_click",
"Box ID": m
}}
```
Another Example:
```json
{{
"Reasoning": "The current screen shows the front page of amazon. There is no previous action. Therefore I need to type "Apple watch" in the search bar.",
"Next Action": "type",
"Box ID": n,
"value": "Apple watch"
}}
```
IMPORTANT NOTES:
1. You should only give a single action at a time.
"""
thinking_model = "r1" in self.model
if not thinking_model:
main_section += """
2. You should give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task.
"""
else:
main_section += """
2. In <think> XML tags give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task. In <output> XML tags put the next action prediction JSON.
"""
main_section += """
3. Attach the next action prediction in the "Next Action".
4. You should not include other actions, such as keyboard shortcuts.
5. When the task is completed, don't complete additional actions. You should say "Next Action": "None" in the json field.
"""
return main_section
def _remove_som_images(messages):
for msg in messages:
msg_content = msg["content"]
if isinstance(msg_content, list):
msg["content"] = [
cnt for cnt in msg_content
if not (isinstance(cnt, str) and 'som' in cnt and is_image_path(cnt))
]
def _maybe_filter_to_n_most_recent_images(
messages: list[BetaMessageParam],
images_to_keep: int,
min_removal_threshold: int = 10,
):
"""
With the assumption that images are screenshots that are of diminishing value as
the conversation progresses, remove all but the final `images_to_keep` tool_result
images in place
"""
if images_to_keep is None:
return messages
total_images = 0
for msg in messages:
for cnt in msg.get("content", []):
if isinstance(cnt, str) and is_image_path(cnt):
total_images += 1
elif isinstance(cnt, dict) and cnt.get("type") == "tool_result":
for content in cnt.get("content", []):
if isinstance(content, dict) and content.get("type") == "image":
total_images += 1
images_to_remove = total_images - images_to_keep
for msg in messages:
msg_content = msg["content"]
if isinstance(msg_content, list):
new_content = []
for cnt in msg_content:
# Remove images from SOM or screenshot as needed
if isinstance(cnt, str) and is_image_path(cnt):
if images_to_remove > 0:
images_to_remove -= 1
continue
# VLM shouldn't use anthropic screenshot tool so shouldn't have these but in case it does, remove as needed
elif isinstance(cnt, dict) and cnt.get("type") == "tool_result":
new_tool_result_content = []
for tool_result_entry in cnt.get("content", []):
if isinstance(tool_result_entry, dict) and tool_result_entry.get("type") == "image":
if images_to_remove > 0:
images_to_remove -= 1
continue
new_tool_result_content.append(tool_result_entry)
cnt["content"] = new_tool_result_content
# Append fixed content to current message's content list
new_content.append(cnt)
msg["content"] = new_content

901
demo/app.py → omnitool/gradio/app.py
View File

@@ -1,479 +1,422 @@
"""
Entrypoint for Gradio, see https://gradio.app/
"""
import platform
import asyncio
import base64
import os
import io
import json
from datetime import datetime
from enum import StrEnum
from functools import partial
from pathlib import Path
from typing import cast, Dict
from PIL import Image
import gradio as gr
from anthropic import APIResponse
from anthropic.types import TextBlock
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
from anthropic.types.tool_use_block import ToolUseBlock
from screeninfo import get_monitors
screens = get_monitors()
print(screens)
from loop import (
PROVIDER_TO_DEFAULT_MODEL_NAME,
APIProvider,
sampling_loop_sync,
)
from tools import ToolResult
from tools.computer import get_screen_details
SCREEN_NAMES, SELECTED_SCREEN_INDEX = get_screen_details()
# SELECTED_SCREEN_INDEX = None
# SCREEN_NAMES = None
CONFIG_DIR = Path("~/.anthropic").expanduser()
API_KEY_FILE = CONFIG_DIR / "api_key"
INTRO_TEXT = '''
🚀🤖✨ It's Play Time!
Welcome to the OmniParser+X Demo! X = [GPT-4o/4o-mini, Claude, Phi, Llama]. Let OmniParser turn your general purpose vision-langauge model to an AI agent. Type a message to play with your beloved assistant.
'''
class Sender(StrEnum):
USER = "user"
BOT = "assistant"
TOOL = "tool"
def setup_state(state):
if "messages" not in state:
state["messages"] = []
if "model" not in state:
# state["model"] = "gpt-4o + ShowUI"
state["model"] = "omniparser + gpt-4o"
# _reset_model(state)
if "provider" not in state:
if state["model"] == "qwen2vl + ShowUI":
state["provider"] = "DashScopeAPI"
elif state["model"] == "gpt-4o + ShowUI":
state["provider"] = "openai"
else:
state["provider"] = os.getenv("API_PROVIDER", "anthropic") or "anthropic"
if "provider_radio" not in state:
state["provider_radio"] = state["provider"]
if "openai_api_key" not in state: # Fetch API keys from environment variables
state["openai_api_key"] = os.getenv("OPENAI_API_KEY", "")
if "anthropic_api_key" not in state:
state["anthropic_api_key"] = os.getenv("ANTHROPIC_API_KEY", "")
if "qwen_api_key" not in state:
state["qwen_api_key"] = os.getenv("QWEN_API_KEY", "")
# Set the initial api_key based on the provider
if "api_key" not in state:
if state["provider"] == "openai":
state["api_key"] = state["openai_api_key"]
elif state["provider"] == "anthropic":
state["api_key"] = state["anthropic_api_key"]
elif state["provider"] == "qwen":
state["api_key"] = state["qwen_api_key"]
else:
state["api_key"] = ""
# print(f"state['api_key']: {state['api_key']}")
if not state["api_key"]:
print("API key not found. Please set it in the environment or paste in textbox.")
if "selected_screen" not in state:
state['selected_screen'] = SELECTED_SCREEN_INDEX if SCREEN_NAMES else 0
if "auth_validated" not in state:
state["auth_validated"] = False
if "responses" not in state:
state["responses"] = {}
if "tools" not in state:
state["tools"] = {}
if "only_n_most_recent_images" not in state:
state["only_n_most_recent_images"] = 10 # 10
if "custom_system_prompt" not in state:
state["custom_system_prompt"] = load_from_storage("system_prompt") or ""
# remove if want to use default system prompt
device_os_name = "Windows" if platform.system() == "Windows" else "Mac" if platform.system() == "Darwin" else "Linux"
state["custom_system_prompt"] += f"\n\nNOTE: you are operating a {device_os_name} machine"
if "hide_images" not in state:
state["hide_images"] = False
if 'chatbot_messages' not in state:
state['chatbot_messages'] = []
def _reset_model(state):
state["model"] = PROVIDER_TO_DEFAULT_MODEL_NAME[cast(APIProvider, state["provider"])]
async def main(state):
"""Render loop for Gradio"""
setup_state(state)
return "Setup completed"
def validate_auth(provider: APIProvider, api_key: str | None):
if provider == APIProvider.ANTHROPIC:
if not api_key:
return "Enter your Anthropic API key to continue."
if provider == APIProvider.BEDROCK:
import boto3
if not boto3.Session().get_credentials():
return "You must have AWS credentials set up to use the Bedrock API."
if provider == APIProvider.VERTEX:
import google.auth
from google.auth.exceptions import DefaultCredentialsError
if not os.environ.get("CLOUD_ML_REGION"):
return "Set the CLOUD_ML_REGION environment variable to use the Vertex API."
try:
google.auth.default(scopes=["https://www.googleapis.com/auth/cloud-platform"])
except DefaultCredentialsError:
return "Your google cloud credentials are not set up correctly."
def load_from_storage(filename: str) -> str | None:
"""Load data from a file in the storage directory."""
try:
file_path = CONFIG_DIR / filename
if file_path.exists():
data = file_path.read_text().strip()
if data:
return data
except Exception as e:
print(f"Debug: Error loading {filename}: {e}")
return None
def save_to_storage(filename: str, data: str) -> None:
"""Save data to a file in the storage directory."""
try:
CONFIG_DIR.mkdir(parents=True, exist_ok=True)
file_path = CONFIG_DIR / filename
file_path.write_text(data)
# Ensure only user can read/write the file
file_path.chmod(0o600)
except Exception as e:
print(f"Debug: Error saving {filename}: {e}")
def _api_response_callback(response: APIResponse[BetaMessage], response_state: dict):
response_id = datetime.now().isoformat()
response_state[response_id] = response
def _tool_output_callback(tool_output: ToolResult, tool_id: str, tool_state: dict):
tool_state[tool_id] = tool_output
def chatbot_output_callback(message, chatbot_state, hide_images=False, sender="bot"):
def _render_message(message: str | BetaTextBlock | BetaToolUseBlock | ToolResult, hide_images=False):
print(f"_render_message: {str(message)[:100]}")
if isinstance(message, str):
return message
is_tool_result = not isinstance(message, str) and (
isinstance(message, ToolResult)
or message.__class__.__name__ == "ToolResult"
or message.__class__.__name__ == "CLIResult"
)
if not message or (
is_tool_result
and hide_images
and not hasattr(message, "error")
and not hasattr(message, "output")
): # return None if hide_images is True
return
# render tool result
if is_tool_result:
message = cast(ToolResult, message)
if message.output:
return message.output
if message.error:
return f"Error: {message.error}"
if message.base64_image and not hide_images:
# somehow can't display via gr.Image
# image_data = base64.b64decode(message.base64_image)
# return gr.Image(value=Image.open(io.BytesIO(image_data)))
return f'<img src="data:image/png;base64,{message.base64_image}">'
elif isinstance(message, BetaTextBlock) or isinstance(message, TextBlock):
return f"Analysis: {message.text}"
elif isinstance(message, BetaToolUseBlock) or isinstance(message, ToolUseBlock):
# return f"Tool Use: {message.name}\nInput: {message.input}"
return f"Next I will perform the following action: {message.input}"
else:
return message
def _truncate_string(s, max_length=500):
"""Truncate long strings for concise printing."""
if isinstance(s, str) and len(s) > max_length:
return s[:max_length] + "..."
return s
# processing Anthropic messages
message = _render_message(message, hide_images)
if sender == "bot":
chatbot_state.append((None, message))
else:
chatbot_state.append((message, None))
# Create a concise version of the chatbot state for printing
concise_state = [(_truncate_string(user_msg), _truncate_string(bot_msg))
for user_msg, bot_msg in chatbot_state]
# print(f"chatbot_output_callback chatbot_state: {concise_state} (truncated)")
def process_input(user_input, state):
setup_state(state)
# Append the user message to state["messages"]
if state["model"] == "gpt-4o + ShowUI" or state["model"] == "qwen2vl + ShowUI":
state["messages"].append(
{
"role": "user",
"content": [TextBlock(type="text", text=user_input)],
}
)
elif state["model"] == "claude-3-5-sonnet-20241022":
state["messages"].append(
{
"role": Sender.USER,
"content": [TextBlock(type="text", text=user_input)],
}
)
elif state["model"] == "omniparser + gpt-4o" or state["model"] == "omniparser + phi35v":
state["messages"].append(
{
"role": "user",
"content": [TextBlock(type="text", text=user_input)],
}
)
# Append the user's message to chatbot_messages with None for the assistant's reply
state['chatbot_messages'].append((user_input, None))
yield state['chatbot_messages'] # Yield to update the chatbot UI with the user's message
# Run sampling_loop_sync with the chatbot_output_callback
for loop_msg in sampling_loop_sync(
system_prompt_suffix=state["custom_system_prompt"],
model=state["model"],
provider=state["provider"],
messages=state["messages"],
output_callback=partial(chatbot_output_callback, chatbot_state=state['chatbot_messages'], hide_images=state["hide_images"]),
tool_output_callback=partial(_tool_output_callback, tool_state=state["tools"]),
api_response_callback=partial(_api_response_callback, response_state=state["responses"]),
api_key=state["api_key"],
only_n_most_recent_images=state["only_n_most_recent_images"],
selected_screen=state['selected_screen']
):
if loop_msg is None:
yield state['chatbot_messages']
print("End of task. Close the loop.")
break
yield state['chatbot_messages'] # Yield the updated chatbot_messages to update the chatbot UI
# with gr.Blocks(theme=gr.themes.Default()) as demo:
with gr.Blocks(theme='YTheme/Minecraft') as demo:
state = gr.State({}) # Use Gradio's state management
setup_state(state.value) # Initialize the state
# Retrieve screen details
gr.Markdown("# OmniParser + ✖️ Demo")
if not os.getenv("HIDE_WARNING", False):
gr.Markdown(INTRO_TEXT)
with gr.Accordion("Settings", open=True):
with gr.Row():
with gr.Column():
model = gr.Dropdown(
label="Model",
choices=["omniparser + gpt-4o", "omniparser + phi35v", "claude-3-5-sonnet-20241022"],
value="omniparser + gpt-4o", # Set to one of the choices
interactive=True,
)
with gr.Column():
provider = gr.Dropdown(
label="API Provider",
choices=[option.value for option in APIProvider],
value="openai",
interactive=False,
)
with gr.Column():
api_key = gr.Textbox(
label="API Key",
type="password",
value=state.value.get("api_key", ""),
placeholder="Paste your API key here",
interactive=True,
)
with gr.Column():
custom_prompt = gr.Textbox(
label="System Prompt Suffix",
value="",
interactive=True,
)
with gr.Column():
screen_options, primary_index = get_screen_details()
SCREEN_NAMES = screen_options
SELECTED_SCREEN_INDEX = primary_index
screen_selector = gr.Dropdown(
label="Select Screen",
choices=screen_options,
value=screen_options[primary_index] if screen_options else None,
interactive=True,
)
with gr.Column():
only_n_images = gr.Slider(
label="N most recent screenshots",
minimum=0,
maximum=10,
step=1,
value=2,
interactive=True,
)
# hide_images = gr.Checkbox(label="Hide screenshots", value=False)
# Define the merged dictionary with task mappings
# merged_dict = json.load(open("examples/ootb_examples.json", "r"))
merged_dict = {}
def update_only_n_images(only_n_images_value, state):
state["only_n_most_recent_images"] = only_n_images_value
# Callback to update the second dropdown based on the first selection
def update_second_menu(selected_category):
return gr.update(choices=list(merged_dict.get(selected_category, {}).keys()))
# Callback to update the third dropdown based on the second selection
def update_third_menu(selected_category, selected_option):
return gr.update(choices=list(merged_dict.get(selected_category, {}).get(selected_option, {}).keys()))
# Callback to update the textbox based on the third selection
def update_textbox(selected_category, selected_option, selected_task):
task_data = merged_dict.get(selected_category, {}).get(selected_option, {}).get(selected_task, {})
prompt = task_data.get("prompt", "")
preview_image = task_data.get("initial_state", "")
task_hint = "Task Hint: " + task_data.get("hint", "")
return prompt, preview_image, task_hint
# Function to update the global variable when the dropdown changes
def update_selected_screen(selected_screen_name, state):
global SCREEN_NAMES
global SELECTED_SCREEN_INDEX
SELECTED_SCREEN_INDEX = SCREEN_NAMES.index(selected_screen_name)
print(f"Selected screen updated to: {SELECTED_SCREEN_INDEX}")
state['selected_screen'] = SELECTED_SCREEN_INDEX
def update_model(model_selection, state):
state["model"] = model_selection
print(f"Model updated to: {state['model']}")
if model_selection == "claude-3-5-sonnet-20241022":
# Provider can be any of the current choices except 'openai'
provider_choices = [option.value for option in APIProvider if option.value != "openai"]
provider_value = "anthropic" # Set default to 'anthropic'
provider_interactive = True
api_key_placeholder = "claude API key"
elif model_selection == "omniparser + gpt-4o" or model_selection == "omniparser + phi35v":
# Provider can be any of the current choices except 'openai'
provider_choices = ["openai"]
provider_value = "openai"
provider_interactive = False
api_key_placeholder = "openai API key"
else:
# Default case
provider_choices = [option.value for option in APIProvider]
provider_value = state.get("provider", provider_choices[0])
provider_interactive = True
api_key_placeholder = ""
# Update the provider in state
state["provider"] = provider_value
# Update api_key in state based on the provider
if provider_value == "openai":
state["api_key"] = state.get("openai_api_key", "")
elif provider_value == "anthropic":
state["api_key"] = state.get("anthropic_api_key", "")
elif provider_value == "qwen":
state["api_key"] = state.get("qwen_api_key", "")
else:
state["api_key"] = ""
# Use gr.update() instead of gr.Dropdown.update()
provider_update = gr.update(
choices=provider_choices,
value=provider_value,
interactive=provider_interactive
)
# Update the API Key textbox
api_key_update = gr.update(
placeholder=api_key_placeholder,
value=state["api_key"]
)
return provider_update, api_key_update
def update_api_key_placeholder(provider_value, model_selection):
if model_selection == "claude-3-5-sonnet-20241022":
if provider_value == "anthropic":
return gr.update(placeholder="anthropic API key")
elif provider_value == "bedrock":
return gr.update(placeholder="bedrock API key")
elif provider_value == "vertex":
return gr.update(placeholder="vertex API key")
else:
return gr.update(placeholder="")
elif model_selection == "gpt-4o + ShowUI":
return gr.update(placeholder="openai API key")
else:
return gr.update(placeholder="")
def update_system_prompt_suffix(system_prompt_suffix, state):
state["custom_system_prompt"] = system_prompt_suffix
api_key.change(fn=lambda key: save_to_storage(API_KEY_FILE, key), inputs=api_key)
with gr.Row():
# submit_button = gr.Button("Submit") # Add submit button
with gr.Column(scale=8):
chat_input = gr.Textbox(show_label=False, placeholder="Type a message to send to Computer Use OOTB...", container=False)
with gr.Column(scale=1, min_width=50):
submit_button = gr.Button(value="Send", variant="primary")
chatbot = gr.Chatbot(label="Chatbot History", autoscroll=True, height=580)
model.change(fn=update_model, inputs=[model, state], outputs=[provider, api_key])
provider.change(fn=update_api_key_placeholder, inputs=[provider, model], outputs=api_key)
screen_selector.change(fn=update_selected_screen, inputs=[screen_selector, state], outputs=None)
only_n_images.change(fn=update_only_n_images, inputs=[only_n_images, state], outputs=None)
# chat_input.submit(process_input, [chat_input, state], chatbot)
submit_button.click(process_input, [chat_input, state], chatbot)
demo.launch(share=True, server_port=7861, server_name='0.0.0.0') # TODO: allowed_paths
"""
python app.py --windows_host_url localhost:8006 --omniparser_server_url localhost:8000
"""
import os
from datetime import datetime
from enum import StrEnum
from functools import partial
from pathlib import Path
from typing import cast
import argparse
import gradio as gr
from anthropic import APIResponse
from anthropic.types import TextBlock
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
from anthropic.types.tool_use_block import ToolUseBlock
from loop import (
APIProvider,
sampling_loop_sync,
)
from tools import ToolResult
import requests
from requests.exceptions import RequestException
import base64
CONFIG_DIR = Path("~/.anthropic").expanduser()
API_KEY_FILE = CONFIG_DIR / "api_key"
INTRO_TEXT = '''
OmniParser lets you turn any vision-langauge model into an AI agent. We currently support OpenAI (4o/o1/o3-mini), DeepSeek (R1), Qwen (2.5VL) or Anthropic Computer Use (Sonnet).
Type a message and press submit to start OmniTool. Press stop to pause, and press the trash icon in the chat to clear the message history.
'''
def parse_arguments():
parser = argparse.ArgumentParser(description="Gradio App")
parser.add_argument("--windows_host_url", type=str, default='localhost:8006')
parser.add_argument("--omniparser_server_url", type=str, default="localhost:8000")
return parser.parse_args()
args = parse_arguments()
class Sender(StrEnum):
USER = "user"
BOT = "assistant"
TOOL = "tool"
def setup_state(state):
if "messages" not in state:
state["messages"] = []
if "model" not in state:
state["model"] = "omniparser + gpt-4o"
if "provider" not in state:
state["provider"] = "openai"
if "openai_api_key" not in state: # Fetch API keys from environment variables
state["openai_api_key"] = os.getenv("OPENAI_API_KEY", "")
if "anthropic_api_key" not in state:
state["anthropic_api_key"] = os.getenv("ANTHROPIC_API_KEY", "")
if "api_key" not in state:
state["api_key"] = ""
if "auth_validated" not in state:
state["auth_validated"] = False
if "responses" not in state:
state["responses"] = {}
if "tools" not in state:
state["tools"] = {}
if "only_n_most_recent_images" not in state:
state["only_n_most_recent_images"] = 2
if 'chatbot_messages' not in state:
state['chatbot_messages'] = []
if 'stop' not in state:
state['stop'] = False
async def main(state):
"""Render loop for Gradio"""
setup_state(state)
return "Setup completed"
def validate_auth(provider: APIProvider, api_key: str | None):
if provider == APIProvider.ANTHROPIC:
if not api_key:
return "Enter your Anthropic API key to continue."
if provider == APIProvider.BEDROCK:
import boto3
if not boto3.Session().get_credentials():
return "You must have AWS credentials set up to use the Bedrock API."
if provider == APIProvider.VERTEX:
import google.auth
from google.auth.exceptions import DefaultCredentialsError
if not os.environ.get("CLOUD_ML_REGION"):
return "Set the CLOUD_ML_REGION environment variable to use the Vertex API."
try:
google.auth.default(scopes=["https://www.googleapis.com/auth/cloud-platform"])
except DefaultCredentialsError:
return "Your google cloud credentials are not set up correctly."
def load_from_storage(filename: str) -> str | None:
"""Load data from a file in the storage directory."""
try:
file_path = CONFIG_DIR / filename
if file_path.exists():
data = file_path.read_text().strip()
if data:
return data
except Exception as e:
print(f"Debug: Error loading {filename}: {e}")
return None
def save_to_storage(filename: str, data: str) -> None:
"""Save data to a file in the storage directory."""
try:
CONFIG_DIR.mkdir(parents=True, exist_ok=True)
file_path = CONFIG_DIR / filename
file_path.write_text(data)
# Ensure only user can read/write the file
file_path.chmod(0o600)
except Exception as e:
print(f"Debug: Error saving {filename}: {e}")
def _api_response_callback(response: APIResponse[BetaMessage], response_state: dict):
response_id = datetime.now().isoformat()
response_state[response_id] = response
def _tool_output_callback(tool_output: ToolResult, tool_id: str, tool_state: dict):
tool_state[tool_id] = tool_output
def chatbot_output_callback(message, chatbot_state, hide_images=False, sender="bot"):
def _render_message(message: str | BetaTextBlock | BetaToolUseBlock | ToolResult, hide_images=False):
print(f"_render_message: {str(message)[:100]}")
if isinstance(message, str):
return message
is_tool_result = not isinstance(message, str) and (
isinstance(message, ToolResult)
or message.__class__.__name__ == "ToolResult"
)
if not message or (
is_tool_result
and hide_images
and not hasattr(message, "error")
and not hasattr(message, "output")
): # return None if hide_images is True
return
# render tool result
if is_tool_result:
message = cast(ToolResult, message)
if message.output:
return message.output
if message.error:
return f"Error: {message.error}"
if message.base64_image and not hide_images:
# somehow can't display via gr.Image
# image_data = base64.b64decode(message.base64_image)
# return gr.Image(value=Image.open(io.BytesIO(image_data)))
return f'<img src="data:image/png;base64,{message.base64_image}">'
elif isinstance(message, BetaTextBlock) or isinstance(message, TextBlock):
return f"Analysis: {message.text}"
elif isinstance(message, BetaToolUseBlock) or isinstance(message, ToolUseBlock):
# return f"Tool Use: {message.name}\nInput: {message.input}"
return f"Next I will perform the following action: {message.input}"
else:
return message
def _truncate_string(s, max_length=500):
"""Truncate long strings for concise printing."""
if isinstance(s, str) and len(s) > max_length:
return s[:max_length] + "..."
return s
# processing Anthropic messages
message = _render_message(message, hide_images)
if sender == "bot":
chatbot_state.append((None, message))
else:
chatbot_state.append((message, None))
# Create a concise version of the chatbot state for printing
concise_state = [(_truncate_string(user_msg), _truncate_string(bot_msg))
for user_msg, bot_msg in chatbot_state]
# print(f"chatbot_output_callback chatbot_state: {concise_state} (truncated)")
def valid_params(user_input, state):
"""Validate all requirements and return a list of error messages."""
errors = []
for server_name, url in [('Windows Host', 'localhost:5000'), ('OmniParser Server', args.omniparser_server_url)]:
try:
url = f'http://{url}/probe'
response = requests.get(url, timeout=3)
if response.status_code != 200:
errors.append(f"{server_name} is not responding")
except RequestException as e:
errors.append(f"{server_name} is not responding")
if not state["api_key"].strip():
errors.append("LLM API Key is not set")
if not user_input:
errors.append("no computer use request provided")
return errors
def process_input(user_input, state):
# Reset the stop flag
if state["stop"]:
state["stop"] = False
errors = valid_params(user_input, state)
if errors:
raise gr.Error("Validation errors: " + ", ".join(errors))
# Append the user message to state["messages"]
state["messages"].append(
{
"role": Sender.USER,
"content": [TextBlock(type="text", text=user_input)],
}
)
# Append the user's message to chatbot_messages with None for the assistant's reply
state['chatbot_messages'].append((user_input, None))
yield state['chatbot_messages'] # Yield to update the chatbot UI with the user's message
print("state")
print(state)
# Run sampling_loop_sync with the chatbot_output_callback
for loop_msg in sampling_loop_sync(
model=state["model"],
provider=state["provider"],
messages=state["messages"],
output_callback=partial(chatbot_output_callback, chatbot_state=state['chatbot_messages'], hide_images=False),
tool_output_callback=partial(_tool_output_callback, tool_state=state["tools"]),
api_response_callback=partial(_api_response_callback, response_state=state["responses"]),
api_key=state["api_key"],
only_n_most_recent_images=state["only_n_most_recent_images"],
max_tokens=16384,
omniparser_url=args.omniparser_server_url
):
if loop_msg is None or state.get("stop"):
yield state['chatbot_messages']
print("End of task. Close the loop.")
break
yield state['chatbot_messages'] # Yield the updated chatbot_messages to update the chatbot UI
def stop_app(state):
state["stop"] = True
return "App stopped"
def get_header_image_base64():
try:
# Get the absolute path to the image relative to this script
script_dir = Path(__file__).parent
image_path = script_dir.parent.parent / "imgs" / "header_bar_thin.png"
with open(image_path, "rb") as image_file:
encoded_string = base64.b64encode(image_file.read()).decode()
return f'data:image/png;base64,{encoded_string}'
except Exception as e:
print(f"Failed to load header image: {e}")
return None
with gr.Blocks(theme=gr.themes.Default()) as demo:
gr.HTML("""
<style>
.no-padding {
padding: 0 !important;
}
.no-padding > div {
padding: 0 !important;
}
</style>
""")
state = gr.State({})
setup_state(state.value)
header_image = get_header_image_base64()
if header_image:
gr.HTML(f'<img src="{header_image}" alt="OmniTool Header" width="100%">', elem_classes="no-padding")
gr.HTML('<h1 style="text-align: center; font-weight: normal;">Omni<span style="font-weight: bold;">Tool</span></h1>')
else:
gr.Markdown("# OmniTool")
if not os.getenv("HIDE_WARNING", False):
gr.Markdown(INTRO_TEXT)
with gr.Accordion("Settings", open=True):
with gr.Row():
with gr.Column():
model = gr.Dropdown(
label="Model",
choices=["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl", "claude-3-5-sonnet-20241022"],
value="omniparser + gpt-4o",
interactive=True,
)
with gr.Column():
only_n_images = gr.Slider(
label="N most recent screenshots",
minimum=0,
maximum=10,
step=1,
value=2,
interactive=True
)
with gr.Row():
with gr.Column(1):
provider = gr.Dropdown(
label="API Provider",
choices=[option.value for option in APIProvider],
value="openai",
interactive=False,
)
with gr.Column(2):
api_key = gr.Textbox(
label="API Key",
type="password",
value=state.value.get("api_key", ""),
placeholder="Paste your API key here",
interactive=True,
)
with gr.Row():
with gr.Column(scale=8):
chat_input = gr.Textbox(show_label=False, placeholder="Type a message to send to Omniparser + X ...", container=False)
with gr.Column(scale=1, min_width=50):
submit_button = gr.Button(value="Send", variant="primary")
with gr.Column(scale=1, min_width=50):
stop_button = gr.Button(value="Stop", variant="secondary")
with gr.Row():
with gr.Column(scale=1):
chatbot = gr.Chatbot(label="Chatbot History", autoscroll=True, height=580)
with gr.Column(scale=3):
iframe = gr.HTML(
f'<iframe src="http://{args.windows_host_url}/vnc.html?view_only=1&autoconnect=1&resize=scale" width="100%" height="580" allow="fullscreen"></iframe>',
container=False,
elem_classes="no-padding"
)
def update_model(model_selection, state):
state["model"] = model_selection
print(f"Model updated to: {state['model']}")
if model_selection == "claude-3-5-sonnet-20241022":
provider_choices = [option.value for option in APIProvider if option.value != "openai"]
elif model_selection in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini"]):
provider_choices = ["openai"]
elif model_selection == "omniparser + R1":
provider_choices = ["groq"]
elif model_selection == "omniparser + qwen2.5vl":
provider_choices = ["dashscope"]
else:
provider_choices = [option.value for option in APIProvider]
default_provider_value = provider_choices[0]
provider_interactive = len(provider_choices) > 1
api_key_placeholder = f"{default_provider_value.title()} API Key"
# Update state
state["provider"] = default_provider_value
state["api_key"] = state.get(f"{default_provider_value}_api_key", "")
# Calls to update other components UI
provider_update = gr.update(
choices=provider_choices,
value=default_provider_value,
interactive=provider_interactive
)
api_key_update = gr.update(
placeholder=api_key_placeholder,
value=state["api_key"]
)
return provider_update, api_key_update
def update_only_n_images(only_n_images_value, state):
state["only_n_most_recent_images"] = only_n_images_value
def update_provider(provider_value, state):
# Update state
state["provider"] = provider_value
state["api_key"] = state.get(f"{provider_value}_api_key", "")
# Calls to update other components UI
api_key_update = gr.update(
placeholder=f"{provider_value.title()} API Key",
value=state["api_key"]
)
return api_key_update
def update_api_key(api_key_value, state):
state["api_key"] = api_key_value
state[f'{state["provider"]}_api_key'] = api_key_value
def clear_chat(state):
# Reset message-related state
state["messages"] = []
state["responses"] = {}
state["tools"] = {}
state['chatbot_messages'] = []
return state['chatbot_messages']
model.change(fn=update_model, inputs=[model, state], outputs=[provider, api_key])
only_n_images.change(fn=update_only_n_images, inputs=[only_n_images, state], outputs=None)
provider.change(fn=update_provider, inputs=[provider, state], outputs=api_key)
api_key.change(fn=update_api_key, inputs=[api_key, state], outputs=None)
chatbot.clear(fn=clear_chat, inputs=[state], outputs=[chatbot])
submit_button.click(process_input, [chat_input, state], chatbot)
stop_button.click(stop_app, [state], None)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7888)

266
demo/executor/anthropic_executor.py → omnitool/gradio/executor/anthropic_executor.py
View File

@@ -1,136 +1,132 @@
import asyncio
from typing import Any, Dict, cast
from collections.abc import Callable
from anthropic.types.beta import (
BetaContentBlock,
BetaContentBlockParam,
BetaImageBlockParam,
BetaMessage,
BetaMessageParam,
BetaTextBlockParam,
BetaToolResultBlockParam,
)
from anthropic.types import TextBlock
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
from tools import BashTool, ComputerTool, EditTool, ToolCollection, ToolResult
class AnthropicExecutor:
def __init__(
self,
output_callback: Callable[[BetaContentBlockParam], None],
tool_output_callback: Callable[[Any, str], None],
selected_screen: int = 0
):
self.tool_collection = ToolCollection(
ComputerTool(selected_screen=selected_screen),
BashTool(),
EditTool(),
)
self.output_callback = output_callback
self.tool_output_callback = tool_output_callback
def __call__(self, response: BetaMessage, messages: list[BetaMessageParam]):
new_message = {
"role": "assistant",
"content": cast(list[BetaContentBlockParam], response.content),
}
if new_message not in messages:
messages.append(new_message)
else:
print("new_message already in messages, there are duplicates.")
tool_result_content: list[BetaToolResultBlockParam] = []
for content_block in cast(list[BetaContentBlock], response.content):
self.output_callback(content_block, sender="bot")
# Execute the tool
if content_block.type == "tool_use":
# Run the asynchronous tool execution in a synchronous context
result = asyncio.run(self.tool_collection.run(
name=content_block.name,
tool_input=cast(dict[str, Any], content_block.input),
))
self.output_callback(result, sender="bot")
tool_result_content.append(
_make_api_tool_result(result, content_block.id)
)
self.tool_output_callback(result, content_block.id)
# Craft messages based on the content_block
# Note: to display the messages in the gradio, you should organize the messages in the following way (user message, bot message)
display_messages = _message_display_callback(messages)
# display_messages = []
# Send the messages to the gradio
for user_msg, bot_msg in display_messages:
# yield [user_msg, bot_msg], tool_result_content
yield [None, None], tool_result_content
if not tool_result_content:
return messages
return tool_result_content
def _message_display_callback(messages):
display_messages = []
for msg in messages:
try:
if isinstance(msg["content"][0], TextBlock):
display_messages.append((msg["content"][0].text, None)) # User message
elif isinstance(msg["content"][0], BetaTextBlock):
display_messages.append((None, msg["content"][0].text)) # Bot message
elif isinstance(msg["content"][0], BetaToolUseBlock):
display_messages.append((None, f"Tool Use: {msg['content'][0].name}\nInput: {msg['content'][0].input}")) # Bot message
elif isinstance(msg["content"][0], Dict) and msg["content"][0]["content"][-1]["type"] == "image":
display_messages.append((None, f'<img src="data:image/png;base64,{msg["content"][0]["content"][-1]["source"]["data"]}">')) # Bot message
else:
print(msg["content"][0])
except Exception as e:
print("error", e)
pass
return display_messages
def _make_api_tool_result(
result: ToolResult, tool_use_id: str
) -> BetaToolResultBlockParam:
"""Convert an agent ToolResult to an API ToolResultBlockParam."""
tool_result_content: list[BetaTextBlockParam | BetaImageBlockParam] | str = []
is_error = False
if result.error:
is_error = True
tool_result_content = _maybe_prepend_system_tool_result(result, result.error)
else:
if result.output:
tool_result_content.append(
{
"type": "text",
"text": _maybe_prepend_system_tool_result(result, result.output),
}
)
if result.base64_image:
tool_result_content.append(
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/png",
"data": result.base64_image,
},
}
)
return {
"type": "tool_result",
"content": tool_result_content,
"tool_use_id": tool_use_id,
"is_error": is_error,
}
def _maybe_prepend_system_tool_result(result: ToolResult, result_text: str):
if result.system:
result_text = f"<system>{result.system}</system>\n{result_text}"
import asyncio
from typing import Any, Dict, cast
from collections.abc import Callable
from anthropic.types.beta import (
BetaContentBlock,
BetaContentBlockParam,
BetaImageBlockParam,
BetaMessage,
BetaMessageParam,
BetaTextBlockParam,
BetaToolResultBlockParam,
)
from anthropic.types import TextBlock
from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
from tools import ComputerTool, ToolCollection, ToolResult
class AnthropicExecutor:
def __init__(
self,
output_callback: Callable[[BetaContentBlockParam], None],
tool_output_callback: Callable[[Any, str], None],
):
self.tool_collection = ToolCollection(
ComputerTool()
)
self.output_callback = output_callback
self.tool_output_callback = tool_output_callback
def __call__(self, response: BetaMessage, messages: list[BetaMessageParam]):
new_message = {
"role": "assistant",
"content": cast(list[BetaContentBlockParam], response.content),
}
if new_message not in messages:
messages.append(new_message)
else:
print("new_message already in messages, there are duplicates.")
tool_result_content: list[BetaToolResultBlockParam] = []
for content_block in cast(list[BetaContentBlock], response.content):
self.output_callback(content_block, sender="bot")
# Execute the tool
if content_block.type == "tool_use":
# Run the asynchronous tool execution in a synchronous context
result = asyncio.run(self.tool_collection.run(
name=content_block.name,
tool_input=cast(dict[str, Any], content_block.input),
))
self.output_callback(result, sender="bot")
tool_result_content.append(
_make_api_tool_result(result, content_block.id)
)
self.tool_output_callback(result, content_block.id)
# Craft messages based on the content_block
# Note: to display the messages in the gradio, you should organize the messages in the following way (user message, bot message)
display_messages = _message_display_callback(messages)
# display_messages = []
# Send the messages to the gradio
for user_msg, bot_msg in display_messages:
# yield [user_msg, bot_msg], tool_result_content
yield [None, None], tool_result_content
if not tool_result_content:
return messages
return tool_result_content
def _message_display_callback(messages):
display_messages = []
for msg in messages:
try:
if isinstance(msg["content"][0], TextBlock):
display_messages.append((msg["content"][0].text, None)) # User message
elif isinstance(msg["content"][0], BetaTextBlock):
display_messages.append((None, msg["content"][0].text)) # Bot message
elif isinstance(msg["content"][0], BetaToolUseBlock):
display_messages.append((None, f"Tool Use: {msg['content'][0].name}\nInput: {msg['content'][0].input}")) # Bot message
elif isinstance(msg["content"][0], Dict) and msg["content"][0]["content"][-1]["type"] == "image":
display_messages.append((None, f'<img src="data:image/png;base64,{msg["content"][0]["content"][-1]["source"]["data"]}">')) # Bot message
else:
print(msg["content"][0])
except Exception as e:
print("error", e)
pass
return display_messages
def _make_api_tool_result(
result: ToolResult, tool_use_id: str
) -> BetaToolResultBlockParam:
"""Convert an agent ToolResult to an API ToolResultBlockParam."""
tool_result_content: list[BetaTextBlockParam | BetaImageBlockParam] | str = []
is_error = False
if result.error:
is_error = True
tool_result_content = _maybe_prepend_system_tool_result(result, result.error)
else:
if result.output:
tool_result_content.append(
{
"type": "text",
"text": _maybe_prepend_system_tool_result(result, result.output),
}
)
if result.base64_image:
tool_result_content.append(
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/png",
"data": result.base64_image,
},
}
)
return {
"type": "tool_result",
"content": tool_result_content,
"tool_use_id": tool_use_id,
"is_error": is_error,
}
def _maybe_prepend_system_tool_result(result: ToolResult, result_text: str):
if result.system:
result_text = f"<system>{result.system}</system>\n{result_text}"
return result_text

114
omnitool/gradio/loop.py Normal file
View File

@@ -0,0 +1,114 @@
"""
Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
"""
from collections.abc import Callable
from enum import StrEnum
from anthropic import APIResponse
from anthropic.types import (
TextBlock,
)
from anthropic.types.beta import (
BetaContentBlock,
BetaMessage,
BetaMessageParam
)
from tools import ToolResult
from agent.llm_utils.omniparserclient import OmniParserClient
from agent.anthropic_agent import AnthropicActor
from agent.vlm_agent import VLMAgent
from executor.anthropic_executor import AnthropicExecutor
BETA_FLAG = "computer-use-2024-10-22"
class APIProvider(StrEnum):
ANTHROPIC = "anthropic"
BEDROCK = "bedrock"
VERTEX = "vertex"
OPENAI = "openai"
PROVIDER_TO_DEFAULT_MODEL_NAME: dict[APIProvider, str] = {
APIProvider.ANTHROPIC: "claude-3-5-sonnet-20241022",
APIProvider.BEDROCK: "anthropic.claude-3-5-sonnet-20241022-v2:0",
APIProvider.VERTEX: "claude-3-5-sonnet-v2@20241022",
APIProvider.OPENAI: "gpt-4o",
}
def sampling_loop_sync(
*,
model: str,
provider: APIProvider | None,
messages: list[BetaMessageParam],
output_callback: Callable[[BetaContentBlock], None],
tool_output_callback: Callable[[ToolResult, str], None],
api_response_callback: Callable[[APIResponse[BetaMessage]], None],
api_key: str,
only_n_most_recent_images: int | None = 2,
max_tokens: int = 4096,
omniparser_url: str
):
"""
Synchronous agentic sampling loop for the assistant/tool interaction of computer use.
"""
print('in sampling_loop_sync, model:', model)
omniparser_client = OmniParserClient(url=f"http://{omniparser_url}/parse/")
if model == "claude-3-5-sonnet-20241022":
# Register Actor and Executor
actor = AnthropicActor(
model=model,
provider=provider,
api_key=api_key,
api_response_callback=api_response_callback,
max_tokens=max_tokens,
only_n_most_recent_images=only_n_most_recent_images
)
elif model in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl"]):
actor = VLMAgent(
model=model,
provider=provider,
api_key=api_key,
api_response_callback=api_response_callback,
output_callback=output_callback,
max_tokens=max_tokens,
only_n_most_recent_images=only_n_most_recent_images
)
else:
raise ValueError(f"Model {model} not supported")
executor = AnthropicExecutor(
output_callback=output_callback,
tool_output_callback=tool_output_callback,
)
print(f"Model Inited: {model}, Provider: {provider}")
tool_result_content = None
print(f"Start the message loop. User messages: {messages}")
if model == "claude-3-5-sonnet-20241022": # Anthropic loop
while True:
parsed_screen = omniparser_client() # parsed_screen: {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, "screen_info"}
screen_info_block = TextBlock(text='Below is the structured accessibility information of the current UI screen, which includes text and icons you can operate on, take these information into account when you are making the prediction for the next action. Note you will still need to take screenshot to get the image: \n' + parsed_screen['screen_info'], type='text')
screen_info_dict = {"role": "user", "content": [screen_info_block]}
messages.append(screen_info_dict)
tools_use_needed = actor(messages=messages)
for message, tool_result_content in executor(tools_use_needed, messages):
yield message
if not tool_result_content:
return messages
messages.append({"content": tool_result_content, "role": "user"})
elif model in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl"]):
while True:
parsed_screen = omniparser_client()
tools_use_needed, vlm_response_json = actor(messages=messages, parsed_screen=parsed_screen)
for message, tool_result_content in executor(tools_use_needed, messages):
yield message
if not tool_result_content:
return messages

27
demo/tools/__init__.py → omnitool/gradio/tools/__init__.py
View File

@@ -1,16 +1,11 @@
from .base import CLIResult, ToolResult
from .bash import BashTool
from .collection import ToolCollection
from .computer import ComputerTool
from .edit import EditTool
from .screen_capture import get_screenshot
__ALL__ = [
BashTool,
CLIResult,
ComputerTool,
EditTool,
ToolCollection,
ToolResult,
get_screenshot,
]
from .base import ToolResult
from .collection import ToolCollection
from .computer import ComputerTool
from .screen_capture import get_screenshot
__ALL__ = [
ComputerTool,
ToolCollection,
ToolResult,
get_screenshot,
]

134
demo/tools/base.py → omnitool/gradio/tools/base.py
View File

@@ -1,69 +1,65 @@
from abc import ABCMeta, abstractmethod
from dataclasses import dataclass, fields, replace
from typing import Any
from anthropic.types.beta import BetaToolUnionParam
class BaseAnthropicTool(metaclass=ABCMeta):
"""Abstract base class for Anthropic-defined tools."""
@abstractmethod
def __call__(self, **kwargs) -> Any:
"""Executes the tool with the given arguments."""
...
@abstractmethod
def to_params(
self,
) -> BetaToolUnionParam:
raise NotImplementedError
@dataclass(kw_only=True, frozen=True)
class ToolResult:
"""Represents the result of a tool execution."""
output: str | None = None
error: str | None = None
base64_image: str | None = None
system: str | None = None
def __bool__(self):
return any(getattr(self, field.name) for field in fields(self))
def __add__(self, other: "ToolResult"):
def combine_fields(
field: str | None, other_field: str | None, concatenate: bool = True
):
if field and other_field:
if concatenate:
return field + other_field
raise ValueError("Cannot combine tool results")
return field or other_field
return ToolResult(
output=combine_fields(self.output, other.output),
error=combine_fields(self.error, other.error),
base64_image=combine_fields(self.base64_image, other.base64_image, False),
system=combine_fields(self.system, other.system),
)
def replace(self, **kwargs):
"""Returns a new ToolResult with the given fields replaced."""
return replace(self, **kwargs)
class CLIResult(ToolResult):
"""A ToolResult that can be rendered as a CLI output."""
class ToolFailure(ToolResult):
"""A ToolResult that represents a failure."""
class ToolError(Exception):
"""Raised when a tool encounters an error."""
def __init__(self, message):
self.message = message
from abc import ABCMeta, abstractmethod
from dataclasses import dataclass, fields, replace
from typing import Any
from anthropic.types.beta import BetaToolUnionParam
class BaseAnthropicTool(metaclass=ABCMeta):
"""Abstract base class for Anthropic-defined tools."""
@abstractmethod
def __call__(self, **kwargs) -> Any:
"""Executes the tool with the given arguments."""
...
@abstractmethod
def to_params(
self,
) -> BetaToolUnionParam:
raise NotImplementedError
@dataclass(kw_only=True, frozen=True)
class ToolResult:
"""Represents the result of a tool execution."""
output: str | None = None
error: str | None = None
base64_image: str | None = None
system: str | None = None
def __bool__(self):
return any(getattr(self, field.name) for field in fields(self))
def __add__(self, other: "ToolResult"):
def combine_fields(
field: str | None, other_field: str | None, concatenate: bool = True
):
if field and other_field:
if concatenate:
return field + other_field
raise ValueError("Cannot combine tool results")
return field or other_field
return ToolResult(
output=combine_fields(self.output, other.output),
error=combine_fields(self.error, other.error),
base64_image=combine_fields(self.base64_image, other.base64_image, False),
system=combine_fields(self.system, other.system),
)
def replace(self, **kwargs):
"""Returns a new ToolResult with the given fields replaced."""
return replace(self, **kwargs)
class ToolFailure(ToolResult):
"""A ToolResult that represents a failure."""
class ToolError(Exception):
"""Raised when a tool encounters an error."""
def __init__(self, message):
self.message = message

68
demo/tools/collection.py → omnitool/gradio/tools/collection.py
View File

@@ -1,34 +1,34 @@
"""Collection classes for managing multiple tools."""
from typing import Any
from anthropic.types.beta import BetaToolUnionParam
from .base import (
BaseAnthropicTool,
ToolError,
ToolFailure,
ToolResult,
)
class ToolCollection:
"""A collection of anthropic-defined tools."""
def __init__(self, *tools: BaseAnthropicTool):
self.tools = tools
self.tool_map = {tool.to_params()["name"]: tool for tool in tools}
def to_params(
self,
) -> list[BetaToolUnionParam]:
return [tool.to_params() for tool in self.tools]
async def run(self, *, name: str, tool_input: dict[str, Any]) -> ToolResult:
tool = self.tool_map.get(name)
if not tool:
return ToolFailure(error=f"Tool {name} is invalid")
try:
return await tool(**tool_input)
except ToolError as e:
return ToolFailure(error=e.message)
"""Collection classes for managing multiple tools."""
from typing import Any
from anthropic.types.beta import BetaToolUnionParam
from .base import (
BaseAnthropicTool,
ToolError,
ToolFailure,
ToolResult,
)
class ToolCollection:
"""A collection of anthropic-defined tools."""
def __init__(self, *tools: BaseAnthropicTool):
self.tools = tools
self.tool_map = {tool.to_params()["name"]: tool for tool in tools}
def to_params(
self,
) -> list[BetaToolUnionParam]:
return [tool.to_params() for tool in self.tools]
async def run(self, *, name: str, tool_input: dict[str, Any]) -> ToolResult:
tool = self.tool_map.get(name)
if not tool:
return ToolFailure(error=f"Tool {name} is invalid")
try:
return await tool(**tool_input)
except ToolError as e:
return ToolFailure(error=e.message)

326
omnitool/gradio/tools/computer.py Normal file
View File

@@ -0,0 +1,326 @@
import base64
import time
from enum import StrEnum
from typing import Literal, TypedDict
from PIL import Image
from anthropic.types.beta import BetaToolComputerUse20241022Param
from .base import BaseAnthropicTool, ToolError, ToolResult
from .screen_capture import get_screenshot
import requests
import re
OUTPUT_DIR = "./tmp/outputs"
TYPING_DELAY_MS = 12
TYPING_GROUP_SIZE = 50
Action = Literal[
"key",
"type",
"mouse_move",
"left_click",
"left_click_drag",
"right_click",
"middle_click",
"double_click",
"screenshot",
"cursor_position",
"hover",
"wait"
]
class Resolution(TypedDict):
width: int
height: int
MAX_SCALING_TARGETS: dict[str, Resolution] = {
"XGA": Resolution(width=1024, height=768), # 4:3
"WXGA": Resolution(width=1280, height=800), # 16:10
"FWXGA": Resolution(width=1366, height=768), # ~16:9
}
class ScalingSource(StrEnum):
COMPUTER = "computer"
API = "api"
class ComputerToolOptions(TypedDict):
display_height_px: int
display_width_px: int
display_number: int | None
def chunks(s: str, chunk_size: int) -> list[str]:
return [s[i : i + chunk_size] for i in range(0, len(s), chunk_size)]
class ComputerTool(BaseAnthropicTool):
"""
A tool that allows the agent to interact with the screen, keyboard, and mouse of the current computer.
Adapted for Windows using 'pyautogui'.
"""
name: Literal["computer"] = "computer"
api_type: Literal["computer_20241022"] = "computer_20241022"
width: int
height: int
display_num: int | None
_screenshot_delay = 2.0
_scaling_enabled = True
@property
def options(self) -> ComputerToolOptions:
width, height = self.scale_coordinates(
ScalingSource.COMPUTER, self.width, self.height
)
return {
"display_width_px": width,
"display_height_px": height,
"display_number": self.display_num,
}
def to_params(self) -> BetaToolComputerUse20241022Param:
return {"name": self.name, "type": self.api_type, **self.options}
def __init__(self, is_scaling: bool = False):
super().__init__()
# Get screen width and height using Windows command
self.display_num = None
self.offset_x = 0
self.offset_y = 0
self.is_scaling = is_scaling
self.width, self.height = self.get_screen_size()
print(f"screen size: {self.width}, {self.height}")
self.key_conversion = {"Page_Down": "pagedown",
"Page_Up": "pageup",
"Super_L": "win",
"Escape": "esc"}
async def __call__(
self,
*,
action: Action,
text: str | None = None,
coordinate: tuple[int, int] | None = None,
**kwargs,
):
print(f"action: {action}, text: {text}, coordinate: {coordinate}, is_scaling: {self.is_scaling}")
if action in ("mouse_move", "left_click_drag"):
if coordinate is None:
raise ToolError(f"coordinate is required for {action}")
if text is not None:
raise ToolError(f"text is not accepted for {action}")
if not isinstance(coordinate, (list, tuple)) or len(coordinate) != 2:
raise ToolError(f"{coordinate} must be a tuple of length 2")
# if not all(isinstance(i, int) and i >= 0 for i in coordinate):
if not all(isinstance(i, int) for i in coordinate):
raise ToolError(f"{coordinate} must be a tuple of non-negative ints")
if self.is_scaling:
x, y = self.scale_coordinates(
ScalingSource.API, coordinate[0], coordinate[1]
)
else:
x, y = coordinate
# print(f"scaled_coordinates: {x}, {y}")
# print(f"offset: {self.offset_x}, {self.offset_y}")
# x += self.offset_x # TODO - check if this is needed
# y += self.offset_y
print(f"mouse move to {x}, {y}")
if action == "mouse_move":
self.send_to_vm(f"pyautogui.moveTo({x}, {y})")
return ToolResult(output=f"Moved mouse to ({x}, {y})")
elif action == "left_click_drag":
current_x, current_y = self.send_to_vm("pyautogui.position()")
self.send_to_vm(f"pyautogui.dragTo({x}, {y}, duration=0.5)")
return ToolResult(output=f"Dragged mouse from ({current_x}, {current_y}) to ({x}, {y})")
if action in ("key", "type"):
if text is None:
raise ToolError(f"text is required for {action}")
if coordinate is not None:
raise ToolError(f"coordinate is not accepted for {action}")
if not isinstance(text, str):
raise ToolError(output=f"{text} must be a string")
if action == "key":
# Handle key combinations
keys = text.split('+')
for key in keys:
key = self.key_conversion.get(key.strip(), key.strip())
key = key.lower()
self.send_to_vm(f"pyautogui.keyDown('{key}')") # Press down each key
for key in reversed(keys):
key = self.key_conversion.get(key.strip(), key.strip())
key = key.lower()
self.send_to_vm(f"pyautogui.keyUp('{key}')") # Release each key in reverse order
return ToolResult(output=f"Pressed keys: {text}")
elif action == "type":
self.send_to_vm(f"pyautogui.typewrite('{text}', interval={TYPING_DELAY_MS / 1000})")
screenshot_base64 = (await self.screenshot()).base64_image
return ToolResult(output=text, base64_image=screenshot_base64)
if action in (
"left_click",
"right_click",
"double_click",
"middle_click",
"screenshot",
"cursor_position",
"left_press",
):
if text is not None:
raise ToolError(f"text is not accepted for {action}")
if coordinate is not None:
raise ToolError(f"coordinate is not accepted for {action}")
if action == "screenshot":
return await self.screenshot()
elif action == "cursor_position":
x, y = self.send_to_vm("pyautogui.position()")
x, y = self.scale_coordinates(ScalingSource.COMPUTER, x, y)
return ToolResult(output=f"X={x},Y={y}")
else:
if action == "left_click":
self.send_to_vm("pyautogui.click()")
elif action == "right_click":
self.send_to_vm("pyautogui.rightClick()")
elif action == "middle_click":
self.send_to_vm("pyautogui.middleClick()")
elif action == "double_click":
self.send_to_vm("pyautogui.doubleClick()")
elif action == "left_press":
self.send_to_vm("pyautogui.mouseDown()")
time.sleep(1)
self.send_to_vm("pyautogui.mouseUp()")
return ToolResult(output=f"Performed {action}")
if action in ("scroll_up", "scroll_down"):
if action == "scroll_up":
self.send_to_vm("pyautogui.scroll(100)")
elif action == "scroll_down":
self.send_to_vm("pyautogui.scroll(-100)")
return ToolResult(output=f"Performed {action}")
if action == "hover":
return ToolResult(output=f"Performed {action}")
if action == "wait":
time.sleep(1)
return ToolResult(output=f"Performed {action}")
raise ToolError(f"Invalid action: {action}")
def send_to_vm(self, action: str):
"""
Executes a python command on the server. Only return tuple of x,y when action is "pyautogui.position()"
"""
prefix = "import pyautogui; pyautogui.FAILSAFE = False;"
command_list = ["python", "-c", f"{prefix} {action}"]
parse = action == "pyautogui.position()"
if parse:
command_list[-1] = f"{prefix} print({action})"
try:
print(f"sending to vm: {command_list}")
response = requests.post(
f"http://localhost:5000/execute",
headers={'Content-Type': 'application/json'},
json={"command": command_list},
timeout=90
)
time.sleep(0.7) # avoid async error as actions take time to complete
print(f"action executed")
if response.status_code != 200:
raise ToolError(f"Failed to execute command. Status code: {response.status_code}")
if parse:
output = response.json()['output'].strip()
match = re.search(r'Point\(x=(\d+),\s*y=(\d+)\)', output)
if not match:
raise ToolError(f"Could not parse coordinates from output: {output}")
x, y = map(int, match.groups())
return x, y
except requests.exceptions.RequestException as e:
raise ToolError(f"An error occurred while trying to execute the command: {str(e)}")
async def screenshot(self):
if not hasattr(self, 'target_dimension'):
screenshot = self.padding_image(screenshot)
self.target_dimension = MAX_SCALING_TARGETS["WXGA"]
width, height = self.target_dimension["width"], self.target_dimension["height"]
screenshot, path = get_screenshot(resize=True, target_width=width, target_height=height)
time.sleep(0.7) # avoid async error as actions take time to complete
return ToolResult(base64_image=base64.b64encode(path.read_bytes()).decode())
def padding_image(self, screenshot):
"""Pad the screenshot to 16:10 aspect ratio, when the aspect ratio is not 16:10."""
_, height = screenshot.size
new_width = height * 16 // 10
padding_image = Image.new("RGB", (new_width, height), (255, 255, 255))
# padding to top left
padding_image.paste(screenshot, (0, 0))
return padding_image
def scale_coordinates(self, source: ScalingSource, x: int, y: int):
"""Scale coordinates to a target maximum resolution."""
if not self._scaling_enabled:
return x, y
ratio = self.width / self.height
target_dimension = None
for target_name, dimension in MAX_SCALING_TARGETS.items():
# allow some error in the aspect ratio - not ratios are exactly 16:9
if abs(dimension["width"] / dimension["height"] - ratio) < 0.02:
if dimension["width"] < self.width:
target_dimension = dimension
self.target_dimension = target_dimension
# print(f"target_dimension: {target_dimension}")
break
if target_dimension is None:
# TODO: currently we force the target to be WXGA (16:10), when it cannot find a match
target_dimension = MAX_SCALING_TARGETS["WXGA"]
self.target_dimension = MAX_SCALING_TARGETS["WXGA"]
# should be less than 1
x_scaling_factor = target_dimension["width"] / self.width
y_scaling_factor = target_dimension["height"] / self.height
if source == ScalingSource.API:
if x > self.width or y > self.height:
raise ToolError(f"Coordinates {x}, {y} are out of bounds")
# scale up
return round(x / x_scaling_factor), round(y / y_scaling_factor)
# scale down
return round(x * x_scaling_factor), round(y * y_scaling_factor)
def get_screen_size(self):
"""Return width and height of the screen"""
try:
response = requests.post(
f"http://localhost:5000/execute",
headers={'Content-Type': 'application/json'},
json={"command": ["python", "-c", "import pyautogui; print(pyautogui.size())"]},
timeout=90
)
if response.status_code != 200:
raise ToolError(f"Failed to get screen size. Status code: {response.status_code}")
output = response.json()['output'].strip()
match = re.search(r'Size\(width=(\d+),\s*height=(\d+)\)', output)
if not match:
raise ToolError(f"Could not parse screen size from output: {output}")
width, height = map(int, match.groups())
return width, height
except requests.exceptions.RequestException as e:
raise ToolError(f"An error occurred while trying to get screen size: {str(e)}")

29
omnitool/gradio/tools/screen_capture.py Normal file
View File

@@ -0,0 +1,29 @@
from pathlib import Path
from uuid import uuid4
import requests
from PIL import Image
from .base import BaseAnthropicTool, ToolError
from io import BytesIO
OUTPUT_DIR = "./tmp/outputs"
def get_screenshot(resize: bool = False, target_width: int = 1920, target_height: int = 1080):
"""Capture screenshot by requesting from HTTP endpoint - returns native resolution unless resized"""
output_dir = Path(OUTPUT_DIR)
output_dir.mkdir(parents=True, exist_ok=True)
path = output_dir / f"screenshot_{uuid4().hex}.png"
try:
response = requests.get('http://localhost:5000/screenshot')
if response.status_code != 200:
raise ToolError(f"Failed to capture screenshot: HTTP {response.status_code}")
# (1280, 800)
screenshot = Image.open(BytesIO(response.content))
if resize and screenshot.size != (target_width, target_height):
screenshot = screenshot.resize((target_width, target_height))
screenshot.save(path)
return screenshot, path
except Exception as e:
raise ToolError(f"Failed to capture screenshot: {str(e)}")

4
omnitool/omnibox/.gitignore vendored Normal file
View File

@@ -0,0 +1,4 @@
vm/win11iso/custom.iso
vm/win11storage
vm/win11setup/setupscripts/firstboot_log.txt
vm/win11setup/setupscripts/server/server.log

48
omnitool/omnibox/Dockerfile Normal file
View File

@@ -0,0 +1,48 @@
ARG VERSION_ARG="latest"
FROM scratch AS build-amd64
COPY --from=qemux/qemu-docker:6.08 / /
ARG DEBCONF_NOWARNINGS="yes"
ARG DEBIAN_FRONTEND="noninteractive"
ARG DEBCONF_NONINTERACTIVE_SEEN="true"
RUN set -eu && \
apt-get update && \
apt-get --no-install-recommends -y install \
bc \
jq \
curl \
7zip \
wsdd \
samba \
xz-utils \
wimtools \
dos2unix \
cabextract \
genisoimage \
libxml2-utils \
libarchive-tools && \
apt-get clean && \
rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
COPY --chmod=755 ./vm/buildcontainer /run/
RUN dos2unix /run/*
COPY --chmod=755 ./vm/win11def /run/assets
RUN dos2unix /run/assets/*
ADD --chmod=755 https://raw.githubusercontent.com/christgau/wsdd/v0.8/src/wsdd.py /usr/sbin/wsdd
ADD --chmod=664 https://github.com/qemus/virtiso-whql/releases/download/v1.9.43-0/virtio-win-1.9.43.tar.xz /drivers.txz
FROM dockurr/windows-arm:${VERSION_ARG} AS build-arm64
FROM build-${TARGETARCH}
ARG VERSION_ARG="0.00"
RUN echo "$VERSION_ARG" > /run/version
EXPOSE 8006 3389
ENV VERSION="win11e"
ENTRYPOINT ["/usr/bin/tini", "-s", "/run/entry.sh"]

23
omnitool/omnibox/compose.yml Normal file
View File

@@ -0,0 +1,23 @@
services:
windows:
image: windows-local
container_name: omni-windows
privileged: true
environment:
RAM_SIZE: "8G"
CPU_CORES: "4"
DISK_SIZE: "20G"
devices:
- /dev/kvm
- /dev/net/tun
cap_add:
- NET_ADMIN
ports:
- 8006:8006 # Web Viewer access
- 5000:5000 # Computer control server
volumes:
- ./vm/win11iso/custom.iso:/custom.iso
- ./vm/win11setup/firstboot:/oem
- ./vm/win11setup/setupscripts:/data
- ./vm/win11storage:/storage

70
omnitool/omnibox/scripts/manage_vm.ps1 Normal file
View File

@@ -0,0 +1,70 @@
function Create-VM {
if (-not (docker images windows-local -q)) {
Write-Host "Image not found locally. Building..."
docker build -t windows-local ..
} else {
Write-Host "Image found locally. Skipping build."
}
docker compose -f ../compose.yml up -d
while ($true) {
try {
$response = Invoke-WebRequest -Uri "http://localhost:5000/probe" -Method GET -UseBasicParsing
if ($response.StatusCode -eq 200) {
break
}
} catch {
Write-Host "Waiting for a response from the computer control server. When first building the VM storage folder this can take a while..."
Start-Sleep -Seconds 5
}
}
Write-Host "VM + server is up and running!"
}
function Start-LocalVM {
Write-Host "Starting VM..."
docker compose -f ../compose.yml start
while ($true) {
try {
$response = Invoke-WebRequest -Uri "http://localhost:5000/probe" -Method GET -UseBasicParsing
if ($response.StatusCode -eq 200) {
break
}
} catch {
Write-Host "Waiting for a response from the computer control server"
Start-Sleep -Seconds 5
}
}
Write-Host "VM started"
}
function Stop-LocalVM {
Write-Host "Stopping VM..."
docker compose -f ../compose.yml stop
Write-Host "VM stopped"
}
function Remove-VM {
Write-Host "Removing VM and associated containers..."
docker compose -f ../compose.yml down
Write-Host "VM removed"
}
if (-not $args[0]) {
Write-Host "Usage: $($MyInvocation.MyCommand.Name) [create|start|stop|delete]"
exit 1
}
switch ($args[0]) {
"create" { Create-VM }
"start" { Start-LocalVM }
"stop" { Stop-LocalVM }
"delete" { Remove-VM }
default {
Write-Host "Invalid option: $($args[0])"
Write-Host "Usage: $($MyInvocation.MyCommand.Name) [create|start|stop|delete]"
exit 1
}
}

77
omnitool/omnibox/scripts/manage_vm.sh Executable file
View File

@@ -0,0 +1,77 @@
#!/bin/bash
create_vm() {
if ! docker images windows-local -q | grep -q .; then
echo "Image not found locally. Building..."
docker build -t windows-local ..
else
echo "Image found locally. Skipping build."
fi
docker compose -f ../compose.yml up -d
# Wait for the VM to start up
while true; do
response=$(curl --write-out '%{http_code}' --silent --output /dev/null localhost:5000/probe)
if [ $response -eq 200 ]; then
break
fi
echo "Waiting for a response from the computer control server. When first building the VM storage folder this can take a while..."
sleep 5
done
echo "VM + server is up and running!"
}
start_vm() {
echo "Starting VM..."
docker compose -f ../compose.yml start
while true; do
response=$(curl --write-out '%{http_code}' --silent --output /dev/null localhost:5000/probe)
if [ $response -eq 200 ]; then
break
fi
echo "Waiting for a response from the computer control server"
sleep 5
done
echo "VM started"
}
stop_vm() {
echo "Stopping VM..."
docker compose -f ../compose.yml stop
echo "VM stopped"
}
delete_vm() {
echo "Removing VM and associated containers..."
docker compose -f ../compose.yml down
echo "VM removed"
}
# Check if control parameter is provided
if [ -z "$1" ]; then
echo "Usage: $0 [create|start|stop|delete]"
exit 1
fi
# Execute the appropriate function based on the control parameter
case "$1" in
"create")
create_vm
;;
"start")
start_vm
;;
"stop")
stop_vm
;;
"delete")
delete_vm
;;
*)
echo "Invalid option: $1"
echo "Usage: $0 [create|start|stop|delete]"
exit 1
;;
esac

410
omnitool/omnibox/vm/buildcontainer/define.sh Normal file
View File

@@ -0,0 +1,410 @@
#!/usr/bin/env bash
set -Eeuo pipefail
: "${WIDTH:=""}"
: "${HEIGHT:=""}"
: "${VERIFY:=""}"
: "${REGION:=""}"
: "${MANUAL:=""}"
: "${REMOVE:=""}"
: "${VERSION:=""}"
: "${DETECTED:=""}"
: "${KEYBOARD:=""}"
: "${LANGUAGE:=""}"
: "${USERNAME:=""}"
: "${PASSWORD:=""}"
MIRRORS=4
PLATFORM="x64"
parseVersion() {
if [[ "${VERSION}" == \"*\" || "${VERSION}" == \'*\' ]]; then
VERSION="${VERSION:1:-1}"
fi
[ -z "$VERSION" ] && VERSION="win11"
case "${VERSION,,}" in
"11" | "11p" | "win11" | "pro11" | "win11p" | "windows11" | "windows 11" )
VERSION="win11x64"
;;
"11e" | "win11e" | "windows11e" | "windows 11e" | "win11x64-enterprise-eval" )
VERSION="win11x64-enterprise-eval"
;;
esac
return 0
}
getLanguage() {
local id="$1"
local ret="$2"
local lang=""
local desc=""
local culture=""
case "${id,,}" in
"ar" | "ar-"* )
lang="Arabic"
desc="$lang"
culture="ar-SA" ;;
"bg" | "bg-"* )
lang="Bulgarian"
desc="$lang"
culture="bg-BG" ;;
"cs" | "cs-"* | "cz" | "cz-"* )
lang="Czech"
desc="$lang"
culture="cs-CZ" ;;
"da" | "da-"* | "dk" | "dk-"* )
lang="Danish"
desc="$lang"
culture="da-DK" ;;
"de" | "de-"* )
lang="German"
desc="$lang"
culture="de-DE" ;;
"el" | "el-"* | "gr" | "gr-"* )
lang="Greek"
desc="$lang"
culture="el-GR" ;;
"gb" | "en-gb" )
lang="English International"
desc="English"
culture="en-GB" ;;
"en" | "en-"* )
lang="English"
desc="English"
culture="en-US" ;;
"mx" | "es-mx" )
lang="Spanish (Mexico)"
desc="Spanish"
culture="es-MX" ;;
"es" | "es-"* )
lang="Spanish"
desc="$lang"
culture="es-ES" ;;
"et" | "et-"* )
lang="Estonian"
desc="$lang"
culture="et-EE" ;;
"fi" | "fi-"* )
lang="Finnish"
desc="$lang"
culture="fi-FI" ;;
"ca" | "fr-ca" )
lang="French Canadian"
desc="French"
culture="fr-CA" ;;
"fr" | "fr-"* )
lang="French"
desc="$lang"
culture="fr-FR" ;;
"he" | "he-"* | "il" | "il-"* )
lang="Hebrew"
desc="$lang"
culture="he-IL" ;;
"hr" | "hr-"* | "cr" | "cr-"* )
lang="Croatian"
desc="$lang"
culture="hr-HR" ;;
"hu" | "hu-"* )
lang="Hungarian"
desc="$lang"
culture="hu-HU" ;;
"it" | "it-"* )
lang="Italian"
desc="$lang"
culture="it-IT" ;;
"ja" | "ja-"* | "jp" | "jp-"* )
lang="Japanese"
desc="$lang"
culture="ja-JP" ;;
"ko" | "ko-"* | "kr" | "kr-"* )
lang="Korean"
desc="$lang"
culture="ko-KR" ;;
"lt" | "lt-"* )
lang="Lithuanian"
desc="$lang"
culture="lv-LV" ;;
"lv" | "lv-"* )
lang="Latvian"
desc="$lang"
culture="lt-LT" ;;
"nb" | "nb-"* |"nn" | "nn-"* | "no" | "no-"* )
lang="Norwegian"
desc="$lang"
culture="nb-NO" ;;
"nl" | "nl-"* )
lang="Dutch"
desc="$lang"
culture="nl-NL" ;;
"pl" | "pl-"* )
lang="Polish"
desc="$lang"
culture="pl-PL" ;;
"br" | "pt-br" )
lang="Brazilian Portuguese"
desc="Portuguese"
culture="pt-BR" ;;
"pt" | "pt-"* )
lang="Portuguese"
desc="$lang"
culture="pt-BR" ;;
"ro" | "ro-"* )
lang="Romanian"
desc="$lang"
culture="ro-RO" ;;
"ru" | "ru-"* )
lang="Russian"
desc="$lang"
culture="ru-RU" ;;
"sk" | "sk-"* )
lang="Slovak"
desc="$lang"
culture="sk-SK" ;;
"sl" | "sl-"* | "si" | "si-"* )
lang="Slovenian"
desc="$lang"
culture="sl-SI" ;;
"sr" | "sr-"* )
lang="Serbian Latin"
desc="Serbian"
culture="sr-Latn-RS" ;;
"sv" | "sv-"* | "se" | "se-"* )
lang="Swedish"
desc="$lang"
culture="sv-SE" ;;
"th" | "th-"* )
lang="Thai"
desc="$lang"
culture="th-TH" ;;
"tr" | "tr-"* )
lang="Turkish"
desc="$lang"
culture="tr-TR" ;;
"ua" | "ua-"* | "uk" | "uk-"* )
lang="Ukrainian"
desc="$lang"
culture="uk-UA" ;;
"hk" | "zh-hk" | "cn-hk" )
lang="Chinese (Traditional)"
desc="Chinese HK"
culture="zh-TW" ;;
"tw" | "zh-tw" | "cn-tw" )
lang="Chinese (Traditional)"
desc="Chinese TW"
culture="zh-TW" ;;
"zh" | "zh-"* | "cn" | "cn-"* )
lang="Chinese (Simplified)"
desc="Chinese"
culture="zh-CN" ;;
esac
case "${ret,,}" in
"desc" ) echo "$desc" ;;
"name" ) echo "$lang" ;;
"culture" ) echo "$culture" ;;
*) echo "$desc";;
esac
return 0
}
parseLanguage() {
REGION="${REGION//_/-/}"
KEYBOARD="${KEYBOARD//_/-/}"
LANGUAGE="${LANGUAGE//_/-/}"
[ -z "$LANGUAGE" ] && LANGUAGE="en"
case "${LANGUAGE,,}" in
"arabic" | "arab" ) LANGUAGE="ar" ;;
"bulgarian" | "bu" ) LANGUAGE="bg" ;;
"chinese" | "cn" ) LANGUAGE="zh" ;;
"croatian" | "cr" | "hrvatski" ) LANGUAGE="hr" ;;
"czech" | "cz" | "cesky" ) LANGUAGE="cs" ;;
"danish" | "dk" | "danske" ) LANGUAGE="da" ;;
"dutch" | "nederlands" ) LANGUAGE="nl" ;;
"english" | "gb" | "british" ) LANGUAGE="en" ;;
"estonian" | "eesti" ) LANGUAGE="et" ;;
"finnish" | "suomi" ) LANGUAGE="fi" ;;
"french" | "français" | "francais" ) LANGUAGE="fr" ;;
"german" | "deutsch" ) LANGUAGE="de" ;;
"greek" | "gr" ) LANGUAGE="el" ;;
"hebrew" | "il" ) LANGUAGE="he" ;;
"hungarian" | "magyar" ) LANGUAGE="hu" ;;
"italian" | "italiano" ) LANGUAGE="it" ;;
"japanese" | "jp" ) LANGUAGE="ja" ;;
"korean" | "kr" ) LANGUAGE="ko" ;;
"latvian" | "latvijas" ) LANGUAGE="lv" ;;
"lithuanian" | "lietuvos" ) LANGUAGE="lt" ;;
"norwegian" | "no" | "nb" | "norsk" ) LANGUAGE="nn" ;;
"polish" | "polski" ) LANGUAGE="pl" ;;
"portuguese" | "pt" | "br" ) LANGUAGE="pt-br" ;;
"português" | "portugues" ) LANGUAGE="pt-br" ;;
"romanian" | "română" | "romana" ) LANGUAGE="ro" ;;
"russian" | "ruski" ) LANGUAGE="ru" ;;
"serbian" | "serbian latin" ) LANGUAGE="sr" ;;
"slovak" | "slovenský" | "slovensky" ) LANGUAGE="sk" ;;
"slovenian" | "si" | "slovenski" ) LANGUAGE="sl" ;;
"spanish" | "espanol" | "español" ) LANGUAGE="es" ;;
"swedish" | "se" | "svenska" ) LANGUAGE="sv" ;;
"turkish" | "türk" | "turk" ) LANGUAGE="tr" ;;
"thai" ) LANGUAGE="th" ;;
"ukrainian" | "ua" ) LANGUAGE="uk" ;;
esac
local culture
culture=$(getLanguage "$LANGUAGE" "culture")
[ -n "$culture" ] && return 0
error "Invalid LANGUAGE specified, value \"$LANGUAGE\" is not recognized!"
return 1
}
printVersion() {
local id="$1"
local desc="$2"
case "${id,,}" in
"win11"* ) desc="Windows 11" ;;
esac
if [ -z "$desc" ]; then
desc="Windows"
[[ "${PLATFORM,,}" != "x64" ]] && desc+=" for ${PLATFORM}"
fi
echo "$desc"
return 0
}
printEdition() {
local id="$1"
local desc="$2"
local result=""
local edition=""
result=$(printVersion "$id" "x")
[[ "$result" == "x" ]] && echo "$desc" && return 0
case "${id,,}" in
*"-enterprise" )
edition="Enterprise"
;;
*"-enterprise-eval" )
edition="Enterprise (Evaluation)"
;;
esac
[ -n "$edition" ] && result+=" $edition"
echo "$result"
return 0
}
fromName() {
local id=""
local name="$1"
local arch="$2"
local add=""
[[ "$arch" != "x64" ]] && add="$arch"
case "${name,,}" in
*"windows 11"* ) id="win11${arch}" ;;
esac
echo "$id"
return 0
}
getVersion() {
local id
local name="$1"
local arch="$2"
id=$(fromName "$name" "$arch")
case "${id,,}" in
"win11"* )
case "${name,,}" in
*" enterprise evaluation"* ) id="$id-enterprise-eval" ;;
*" enterprise"* ) id="$id-enterprise" ;;
esac
;;
esac
echo "$id"
return 0
}
addFolder() {
local src="$1"
local folder="/oem"
[ ! -d "$folder" ] && folder="/OEM"
[ ! -d "$folder" ] && folder="$STORAGE/oem"
[ ! -d "$folder" ] && folder="$STORAGE/OEM"
[ ! -d "$folder" ] && return 0
local msg="Adding OEM folder to image..."
info "$msg" && html "$msg"
local dest="$src/\$OEM\$/\$1/OEM"
mkdir -p "$dest" || return 1
cp -Lr "$folder/." "$dest" || return 1
local file
file=$(find "$dest" -maxdepth 1 -type f -iname install.bat | head -n 1)
[ -f "$file" ] && unix2dos -q "$file"
return 0
}
# migrateFiles() {
# local base="$1"
# local version="$2"
# local file=""
# [ -f "$base" ] && return 0
# [[ "${version,,}" == "tiny10" ]] && file="tiny10_x64_23h2.iso"
# [[ "${version,,}" == "tiny11" ]] && file="tiny11_2311_x64.iso"
# [[ "${version,,}" == "core11" ]] && file="tiny11_core_x64_beta_1.iso"
# [[ "${version,,}" == "winxpx86" ]] && file="en_windows_xp_professional_with_service_pack_3_x86_cd_x14-80428.iso"
# [[ "${version,,}" == "winvistax64" ]] && file="en_windows_vista_sp2_x64_dvd_342267.iso"
# [[ "${version,,}" == "win7x64" ]] && file="en_windows_7_enterprise_with_sp1_x64_dvd_u_677651.iso"
# [ ! -f "$STORAGE/$file" ] && return 0
# mv -f "$STORAGE/$file" "$base" || return 1
# return 0
# }
migrateFiles() {
local base="$1"
local version="$2"
local file=""
[ -f "$base" ] && return 0
[ ! -f "$STORAGE/$file" ] && return 0
mv -f "$STORAGE/$file" "$base" || return 1
return 0
}
return 0

38
omnitool/omnibox/vm/buildcontainer/entry.sh Normal file
View File

@@ -0,0 +1,38 @@
#!/usr/bin/env bash
set -Eeuo pipefail
: "${BOOT_MODE:="windows"}"
APP="OmniParser Windows"
SUPPORT="https://github.com/microsoft/OmniParser"
cd /run
. reset.sh # Initialize system
. define.sh # Define versions
. install.sh # Run installation
. disk.sh # Initialize disks
. display.sh # Initialize graphics
. network.sh # Initialize network
. samba.sh # Configure samba
. boot.sh # Configure boot
. proc.sh # Initialize processor
. power.sh # Configure shutdown
. config.sh # Configure arguments
trap - ERR
version=$(qemu-system-x86_64 --version | head -n 1 | cut -d '(' -f 1 | awk '{ print $NF }')
info "Booting ${APP}${BOOT_DESC} using QEMU v$version..."
{ qemu-system-x86_64 ${ARGS:+ $ARGS} >"$QEMU_OUT" 2>"$QEMU_LOG"; rc=$?; } || :
(( rc != 0 )) && error "$(<"$QEMU_LOG")" && exit 15
terminal
( sleep 30; boot ) &
tail -fn +0 "$QEMU_LOG" 2>/dev/null &
cat "$QEMU_TERM" 2> /dev/null | tee "$QEMU_PTY" &
wait $! || :
sleep 1 & wait $!
[ ! -f "$QEMU_END" ] && finish 0

1024
omnitool/omnibox/vm/buildcontainer/install.sh Normal file
View File

File diff suppressed because it is too large Load Diff

223
omnitool/omnibox/vm/buildcontainer/power.sh Normal file
View File

@@ -0,0 +1,223 @@
#!/usr/bin/env bash
set -Eeuo pipefail
# Configure QEMU for graceful shutdown
QEMU_TERM=""
QEMU_PORT=7100
QEMU_TIMEOUT=110
QEMU_DIR="/run/shm"
QEMU_PID="$QEMU_DIR/qemu.pid"
QEMU_PTY="$QEMU_DIR/qemu.pty"
QEMU_LOG="$QEMU_DIR/qemu.log"
QEMU_OUT="$QEMU_DIR/qemu.out"
QEMU_END="$QEMU_DIR/qemu.end"
rm -f "$QEMU_DIR/qemu.*"
touch "$QEMU_LOG"
_trap() {
func="$1" ; shift
for sig ; do
trap "$func $sig" "$sig"
done
}
boot() {
[ -f "$QEMU_END" ] && return 0
if [ -s "$QEMU_PTY" ]; then
if [ "$(stat -c%s "$QEMU_PTY")" -gt 7 ]; then
local fail=""
if [[ "${BOOT_MODE,,}" == "windows_legacy" ]]; then
grep -Fq "No bootable device." "$QEMU_PTY" && fail="y"
grep -Fq "BOOTMGR is missing" "$QEMU_PTY" && fail="y"
fi
if [ -z "$fail" ]; then
info "Windows has started successfully. You can directly view the VM at http://localhost:8006/vnc.html?view_only=1&autoconnect=1&resize=scale. Wait until setup is complete before interacting manually."
return 0
fi
fi
fi
error "Timeout while waiting for QEMU to boot the machine!"
local pid
pid=$(<"$QEMU_PID")
{ kill -15 "$pid" || true; } 2>/dev/null
return 0
}
ready() {
[ -f "$STORAGE/windows.boot" ] && return 0
[ ! -s "$QEMU_PTY" ] && return 1
if [[ "${BOOT_MODE,,}" == "windows_legacy" ]]; then
local last
local bios="Booting from Hard"
last=$(grep "^Booting.*" "$QEMU_PTY" | tail -1)
[[ "${last,,}" != "${bios,,}"* ]] && return 1
grep -Fq "No bootable device." "$QEMU_PTY" && return 1
grep -Fq "BOOTMGR is missing" "$QEMU_PTY" && return 1
return 0
fi
local line="\"Windows Boot Manager\""
grep -Fq "$line" "$QEMU_PTY" && return 0
return 1
}
finish() {
local pid
local reason=$1
touch "$QEMU_END"
if [ -s "$QEMU_PID" ]; then
pid=$(<"$QEMU_PID")
error "Forcefully terminating Windows, reason: $reason..."
{ kill -15 "$pid" || true; } 2>/dev/null
while isAlive "$pid"; do
sleep 1
# Workaround for zombie pid
[ ! -s "$QEMU_PID" ] && break
done
fi
if [ ! -f "$STORAGE/windows.boot" ] && [ -f "$BOOT" ]; then
# Remove CD-ROM ISO after install
if ready; then
touch "$STORAGE/windows.boot"
if [[ "$REMOVE" != [Nn]* ]]; then
rm -f "$BOOT" 2>/dev/null || true
fi
fi
fi
pid="/var/run/tpm.pid"
[ -s "$pid" ] && pKill "$(<"$pid")"
pid="/var/run/wsdd.pid"
[ -s "$pid" ] && pKill "$(<"$pid")"
fKill "smbd"
closeNetwork
sleep 0.5
echo "❯ Shutdown completed!"
exit "$reason"
}
terminal() {
local dev=""
if [ -s "$QEMU_OUT" ]; then
local msg
msg=$(<"$QEMU_OUT")
if [ -n "$msg" ]; then
if [[ "${msg,,}" != "char"* || "$msg" != *"serial0)" ]]; then
echo "$msg"
fi
dev="${msg#*/dev/p}"
dev="/dev/p${dev%% *}"
fi
fi
if [ ! -c "$dev" ]; then
dev=$(echo 'info chardev' | nc -q 1 -w 1 localhost "$QEMU_PORT" | tr -d '\000')
dev="${dev#*serial0}"
dev="${dev#*pty:}"
dev="${dev%%$'\n'*}"
dev="${dev%%$'\r'*}"
fi
if [ ! -c "$dev" ]; then
error "Device '$dev' not found!"
finish 34 && return 34
fi
QEMU_TERM="$dev"
return 0
}
_graceful_shutdown() {
local code=$?
set +e
if [ -f "$QEMU_END" ]; then
info "Received $1 while already shutting down..."
return
fi
touch "$QEMU_END"
info "Received $1, sending ACPI shutdown signal..."
if [ ! -s "$QEMU_PID" ]; then
error "QEMU PID file does not exist?"
finish "$code" && return "$code"
fi
local pid=""
pid=$(<"$QEMU_PID")
if ! isAlive "$pid"; then
error "QEMU process does not exist?"
finish "$code" && return "$code"
fi
if ! ready; then
info "Cannot send ACPI signal during Windows setup, aborting..."
finish "$code" && return "$code"
fi
# Send ACPI shutdown signal
echo 'system_powerdown' | nc -q 1 -w 1 localhost "${QEMU_PORT}" > /dev/null
local cnt=0
while [ "$cnt" -lt "$QEMU_TIMEOUT" ]; do
sleep 1
cnt=$((cnt+1))
! isAlive "$pid" && break
# Workaround for zombie pid
[ ! -s "$QEMU_PID" ] && break
info "Waiting for Windows to shutdown... ($cnt/$QEMU_TIMEOUT)"
# Send ACPI shutdown signal
echo 'system_powerdown' | nc -q 1 -w 1 localhost "${QEMU_PORT}" > /dev/null
done
if [ "$cnt" -ge "$QEMU_TIMEOUT" ]; then
error "Shutdown timeout reached, aborting..."
fi
finish "$code" && return "$code"
}
SERIAL="pty"
MONITOR="telnet:localhost:$QEMU_PORT,server,nowait,nodelay"
MONITOR+=" -daemonize -D $QEMU_LOG -pidfile $QEMU_PID"
_trap _graceful_shutdown SIGTERM SIGHUP SIGINT SIGABRT SIGQUIT
return 0

109
omnitool/omnibox/vm/buildcontainer/samba.sh Normal file
View File

@@ -0,0 +1,109 @@
#!/usr/bin/env bash
set -Eeuo pipefail
: "${SAMBA:="Y"}"
[[ "$SAMBA" == [Nn]* ]] && return 0
[[ "$NETWORK" == [Nn]* ]] && return 0
hostname="host.lan"
interface="dockerbridge"
if [[ "$DHCP" == [Yy1]* ]]; then
hostname="$IP"
interface="$VM_NET_DEV"
fi
addShare() {
local dir="$1"
local name="$2"
local comment="$3"
mkdir -p "$dir" || return 1
if [ -z "$(ls -A "$dir")" ]; then
chmod 777 "$dir"
{ echo "--------------------------------------------------------"
echo " $APP"
echo " For support visit $SUPPORT"
echo "--------------------------------------------------------"
echo ""
echo "Using this folder you can share files with the host machine."
echo ""
echo "To change its location, include the following bind mount in your compose file:"
echo ""
echo " volumes:"
echo " - \"/home/example:/${name,,}\""
echo ""
echo "Or in your run command:"
echo ""
echo " -v \"/home/example:/${name,,}\""
echo ""
echo "Replace the example path /home/example with the desired shared folder."
echo ""
} | unix2dos > "$dir/readme.txt"
fi
{ echo ""
echo "[$name]"
echo " path = $dir"
echo " comment = $comment"
echo " writable = yes"
echo " guest ok = yes"
echo " guest only = yes"
echo " force user = root"
echo " force group = root"
} >> "/etc/samba/smb.conf"
return 0
}
{ echo "[global]"
echo " server string = Dockur"
echo " netbios name = $hostname"
echo " workgroup = WORKGROUP"
echo " interfaces = $interface"
echo " bind interfaces only = yes"
echo " security = user"
echo " guest account = nobody"
echo " map to guest = Bad User"
echo " server min protocol = NT1"
echo ""
echo " # disable printing services"
echo " load printers = no"
echo " printing = bsd"
echo " printcap name = /dev/null"
echo " disable spoolss = yes"
} > "/etc/samba/smb.conf"
share="/data"
[ ! -d "$share" ] && [ -d "$STORAGE/data" ] && share="$STORAGE/data"
[ ! -d "$share" ] && [ -d "/shared" ] && share="/shared"
[ ! -d "$share" ] && [ -d "$STORAGE/shared" ] && share="$STORAGE/shared"
addShare "$share" "Data" "Shared" || error "Failed to create shared folder!"
[ -d "/data2" ] && addShare "/data2" "Data2" "Shared"
[ -d "/data3" ] && addShare "/data3" "Data3" "Shared"
if ! smbd; then
error "Samba daemon failed to start!"
smbd -i --debug-stdout || true
fi
if [[ "${BOOT_MODE:-}" == "windows_legacy" ]]; then
# Enable NetBIOS on Windows 7 and lower
if ! nmbd; then
error "NetBIOS daemon failed to start!"
nmbd -i --debug-stdout || true
fi
else
# Enable Web Service Discovery on Vista and up
wsdd -i "$interface" -p -n "$hostname" &
echo "$!" > /var/run/wsdd.pid
fi
return 0

462
omnitool/omnibox/vm/win11def/win11x64-enterprise-eval.xml Normal file
View File

@@ -0,0 +1,462 @@
<?xml version="1.0" encoding="UTF-8"?>
<unattend xmlns="urn:schemas-microsoft-com:unattend" xmlns:wcm="http://schemas.microsoft.com/WMIConfig/2002/State">
<settings pass="windowsPE">
<component name="Microsoft-Windows-International-Core-WinPE" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<SetupUILanguage>
<UILanguage>en-US</UILanguage>
</SetupUILanguage>
<InputLocale>0409:00000409</InputLocale>
<SystemLocale>en-US</SystemLocale>
<UILanguage>en-US</UILanguage>
<UserLocale>en-US</UserLocale>
</component>
<component name="Microsoft-Windows-Setup" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<DiskConfiguration>
<Disk wcm:action="add">
<DiskID>0</DiskID>
<WillWipeDisk>true</WillWipeDisk>
<CreatePartitions>
<!-- System partition (ESP) -->
<CreatePartition wcm:action="add">
<Order>1</Order>
<Type>EFI</Type>
<Size>128</Size>
</CreatePartition>
<!-- Microsoft reserved partition (MSR) -->
<CreatePartition wcm:action="add">
<Order>2</Order>
<Type>MSR</Type>
<Size>128</Size>
</CreatePartition>
<!-- Windows partition -->
<CreatePartition wcm:action="add">
<Order>3</Order>
<Type>Primary</Type>
<Extend>true</Extend>
</CreatePartition>
</CreatePartitions>
<ModifyPartitions>
<!-- System partition (ESP) -->
<ModifyPartition wcm:action="add">
<Order>1</Order>
<PartitionID>1</PartitionID>
<Label>System</Label>
<Format>FAT32</Format>
</ModifyPartition>
<!-- MSR partition does not need to be modified -->
<ModifyPartition wcm:action="add">
<Order>2</Order>
<PartitionID>2</PartitionID>
</ModifyPartition>
<!-- Windows partition -->
<ModifyPartition wcm:action="add">
<Order>3</Order>
<PartitionID>3</PartitionID>
<Label>Windows</Label>
<Letter>C</Letter>
<Format>NTFS</Format>
</ModifyPartition>
</ModifyPartitions>
</Disk>
</DiskConfiguration>
<ImageInstall>
<OSImage>
<InstallTo>
<DiskID>0</DiskID>
<PartitionID>3</PartitionID>
</InstallTo>
<InstallToAvailablePartition>false</InstallToAvailablePartition>
</OSImage>
</ImageInstall>
<DynamicUpdate>
<Enable>true</Enable>
<WillShowUI>Never</WillShowUI>
</DynamicUpdate>
<UpgradeData>
<Upgrade>false</Upgrade>
<WillShowUI>Never</WillShowUI>
</UpgradeData>
<UserData>
<AcceptEula>true</AcceptEula>
<FullName>Docker</FullName>
<Organization>Windows for Docker</Organization>
</UserData>
<EnableFirewall>false</EnableFirewall>
<Diagnostics>
<OptIn>false</OptIn>
</Diagnostics>
<RunSynchronous>
<RunSynchronousCommand wcm:action="add">
<Order>1</Order>
<Path>reg.exe add "HKLM\SYSTEM\Setup\LabConfig" /v BypassTPMCheck /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>2</Order>
<Path>reg.exe add "HKLM\SYSTEM\Setup\LabConfig" /v BypassSecureBootCheck /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>3</Order>
<Path>reg.exe add "HKLM\SYSTEM\Setup\LabConfig" /v BypassRAMCheck /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>4</Order>
<Path>reg.exe add "HKLM\SYSTEM\Setup\MoSetup" /v AllowUpgradesWithUnsupportedTPMOrCPU /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
</RunSynchronous>
</component>
</settings>
<settings pass="offlineServicing">
<component name="Microsoft-Windows-LUA-Settings" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<EnableLUA>false</EnableLUA>
</component>
</settings>
<settings pass="generalize">
<component name="Microsoft-Windows-PnPSysprep" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<PersistAllDeviceInstalls>true</PersistAllDeviceInstalls>
</component>
<component name="Microsoft-Windows-Security-SPP" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<SkipRearm>1</SkipRearm>
</component>
</settings>
<settings pass="specialize">
<component name="Microsoft-Windows-Security-SPP-UX" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<SkipAutoActivation>true</SkipAutoActivation>
</component>
<component name="Microsoft-Windows-Shell-Setup" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<ComputerName>*</ComputerName>
<OEMInformation>
<Manufacturer>Dockur</Manufacturer>
<Model>Windows for Docker</Model>
<SupportHours>24/7</SupportHours>
<SupportPhone />
<SupportProvider>Dockur</SupportProvider>
<SupportURL>https://github.com/dockur/windows/issues</SupportURL>
</OEMInformation>
<OEMName>Windows for Docker</OEMName>
</component>
<component name="Microsoft-Windows-ErrorReportingCore" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<DisableWER>1</DisableWER>
</component>
<component name="Microsoft-Windows-IE-InternetExplorer" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<DisableAccelerators>true</DisableAccelerators>
<DisableFirstRunWizard>true</DisableFirstRunWizard>
<Home_Page>https://google.com</Home_Page>
<Help_Page>about:blank</Help_Page>
</component>
<component name="Microsoft-Windows-IE-InternetExplorer" processorArchitecture="wow64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<DisableAccelerators>true</DisableAccelerators>
<DisableFirstRunWizard>true</DisableFirstRunWizard>
<Home_Page>https://google.com</Home_Page>
<Help_Page>about:blank</Help_Page>
</component>
<component name="Microsoft-Windows-SQMApi" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<CEIPEnabled>0</CEIPEnabled>
</component>
<component name="Microsoft-Windows-SystemRestore-Main" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<DisableSR>1</DisableSR>
</component>
<component name="Microsoft-Windows-International-Core" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<InputLocale>0409:00000409</InputLocale>
<SystemLocale>en-US</SystemLocale>
<UILanguage>en-US</UILanguage>
<UserLocale>en-US</UserLocale>
</component>
<component name="Microsoft-Windows-Deployment" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<RunSynchronous>
<RunSynchronousCommand wcm:action="add">
<Order>1</Order>
<Path>reg.exe add "HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\OOBE" /v BypassNRO /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>2</Order>
<Path>reg.exe load "HKU\mount" "C:\Users\Default\NTUSER.DAT"</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>3</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "ContentDeliveryAllowed" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>4</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "FeatureManagementEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>5</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "OEMPreInstalledAppsEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>6</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "PreInstalledAppsEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>7</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "PreInstalledAppsEverEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>8</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SilentInstalledAppsEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>9</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SoftLandingEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>10</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContentEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>11</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContent-310093Enabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>12</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContent-338387Enabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>13</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContent-338388Enabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>14</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContent-338389Enabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>15</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContent-338393Enabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>16</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SubscribedContent-353698Enabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>17</Order>
<Path>reg.exe add "HKU\mount\Software\Microsoft\Windows\CurrentVersion\ContentDeliveryManager" /v "SystemPaneSuggestionsEnabled" /t REG_DWORD /d 0 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>18</Order>
<Path>reg.exe add "HKU\mount\Software\Policies\Microsoft\Windows\CloudContent" /v "DisableCloudOptimizedContent" /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>19</Order>
<Path>reg.exe add "HKU\mount\Software\Policies\Microsoft\Windows\CloudContent" /v "DisableWindowsConsumerFeatures" /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>20</Order>
<Path>reg.exe add "HKU\mount\Software\Policies\Microsoft\Windows\CloudContent" /v "DisableConsumerAccountStateContent" /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>21</Order>
<Path>reg.exe unload "HKU\mount"</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>22</Order>
<Path>reg.exe add "HKLM\Software\Policies\Microsoft\Windows\CloudContent" /v "DisableCloudOptimizedContent" /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>23</Order>
<Path>reg.exe add "HKLM\Software\Policies\Microsoft\Windows\CloudContent" /v "DisableWindowsConsumerFeatures" /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>24</Order>
<Path>reg.exe add "HKLM\Software\Policies\Microsoft\Windows\CloudContent" /v "DisableConsumerAccountStateContent" /t REG_DWORD /d 1 /f</Path>
</RunSynchronousCommand>
<RunSynchronousCommand wcm:action="add">
<Order>25</Order>
<Path>reg.exe add "HKLM\SOFTWARE\Policies\Microsoft\Windows NT\CurrentVersion\NetworkList\Signatures\FirstNetwork" /v Category /t REG_DWORD /d 1 /f</Path>
<Description>Set Network Location to Home</Description>
</RunSynchronousCommand>
</RunSynchronous>
</component>
<component name="Microsoft-Windows-TerminalServices-LocalSessionManager" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<fDenyTSConnections>false</fDenyTSConnections>
</component>
<component name="Microsoft-Windows-TerminalServices-RDP-WinStationExtensions" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<UserAuthentication>0</UserAuthentication>
</component>
<component name="Networking-MPSSVC-Svc" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<FirewallGroups>
<FirewallGroup wcm:action="add" wcm:keyValue="RemoteDesktop">
<Active>true</Active>
<Profile>all</Profile>
<Group>@FirewallAPI.dll,-28752</Group>
</FirewallGroup>
</FirewallGroups>
</component>
</settings>
<settings pass="auditSystem" />
<settings pass="auditUser" />
<settings pass="oobeSystem">
<component name="Microsoft-Windows-SecureStartup-FilterDriver" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<PreventDeviceEncryption>true</PreventDeviceEncryption>
</component>
<component name="Microsoft-Windows-EnhancedStorage-Adm" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<TCGSecurityActivationDisabled>1</TCGSecurityActivationDisabled>
</component>
<component name="Microsoft-Windows-Shell-Setup" processorArchitecture="amd64" publicKeyToken="31bf3856ad364e35" language="neutral" versionScope="nonSxS">
<UserAccounts>
<LocalAccounts>
<LocalAccount wcm:action="add">
<Name>Docker</Name>
<Group>Administrators</Group>
<Password>
<Value />
<PlainText>true</PlainText>
</Password>
</LocalAccount>
</LocalAccounts>
<AdministratorPassword>
<Value>password</Value>
<PlainText>true</PlainText>
</AdministratorPassword>
</UserAccounts>
<AutoLogon>
<Username>Docker</Username>
<Enabled>true</Enabled>
<LogonCount>65432</LogonCount>
<Password>
<Value />
<PlainText>true</PlainText>
</Password>
</AutoLogon>
<Display>
<ColorDepth>32</ColorDepth>
<HorizontalResolution>1920</HorizontalResolution>
<VerticalResolution>1080</VerticalResolution>
</Display>
<OOBE>
<HideEULAPage>true</HideEULAPage>
<HideLocalAccountScreen>true</HideLocalAccountScreen>
<HideOEMRegistrationScreen>true</HideOEMRegistrationScreen>
<HideOnlineAccountScreens>true</HideOnlineAccountScreens>
<HideWirelessSetupInOOBE>true</HideWirelessSetupInOOBE>
<NetworkLocation>Home</NetworkLocation>
<ProtectYourPC>3</ProtectYourPC>
<SkipUserOOBE>true</SkipUserOOBE>
<SkipMachineOOBE>true</SkipMachineOOBE>
</OOBE>
<RegisteredOrganization>Dockur</RegisteredOrganization>
<RegisteredOwner>Windows for Docker</RegisteredOwner>
<FirstLogonCommands>
<SynchronousCommand wcm:action="add">
<Order>1</Order>
<CommandLine>reg.exe add "HKLM\SYSTEM\CurrentControlSet\Services\LanmanWorkstation\Parameters" /v "AllowInsecureGuestAuth" /t REG_DWORD /d 1 /f</CommandLine>
<Description>Allow guest access to network shares</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>2</Order>
<CommandLine>reg.exe add "HKLM\SYSTEM\CurrentControlSet\Services\LanmanWorkstation\Parameters" /v "RequireSecuritySignature" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Disable SMB signing requirement</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>3</Order>
<CommandLine>reg.exe add "HKLM\SYSTEM\CurrentControlSet\Control\Lsa" /v LimitBlankPasswordUse /t REG_DWORD /d 0 /f</CommandLine>
<Description>Allow RDP login with blank password</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>4</Order>
<CommandLine>reg.exe add "HKLM\SOFTWARE\Microsoft\Windows NT\CurrentVersion\PasswordLess\Device" /v "DevicePasswordLessBuildVersion" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Enable option for passwordless sign-in</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>5</Order>
<CommandLine>cmd /C wmic useraccount where name="Docker" set PasswordExpires=false</CommandLine>
<Description>Password Never Expires</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>6</Order>
<CommandLine>cmd /C POWERCFG -H OFF</CommandLine>
<Description>Disable Hibernation</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>7</Order>
<CommandLine>cmd /C POWERCFG -X -monitor-timeout-ac 0</CommandLine>
<Description>Disable monitor blanking</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>8</Order>
<CommandLine>reg.exe add "HKLM\SOFTWARE\Policies\Microsoft\Edge" /v "HideFirstRunExperience" /t REG_DWORD /d 1 /f</CommandLine>
<Description>Disable first-run experience in Edge</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>9</Order>
<CommandLine>reg.exe add "HKCU\SOFTWARE\Microsoft\Windows\CurrentVersion\Explorer\Advanced" /v "HideFileExt" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Show file extensions in Explorer</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>10</Order>
<CommandLine>reg.exe add "HKLM\SYSTEM\CurrentControlSet\Control\Power" /v "HibernateFileSizePercent" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Zero Hibernation File</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>11</Order>
<CommandLine>reg.exe add "HKLM\SYSTEM\CurrentControlSet\Control\Power" /v "HibernateEnabled" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Disable Hibernation</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>12</Order>
<CommandLine>cmd /C POWERCFG -X -standby-timeout-ac 0</CommandLine>
<Description>Disable Sleep</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>13</Order>
<CommandLine>reg.exe add "HKLM\SOFTWARE\Policies\Microsoft\Windows NT\Terminal Services" /v "fAllowUnlistedRemotePrograms" /t REG_DWORD /d 1 /f</CommandLine>
<Description>Enable RemoteAPP to launch unlisted programs</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>14</Order>
<CommandLine>reg.exe add "HKCU\SOFTWARE\Microsoft\Windows\CurrentVersion\Explorer\Advanced" /v "ShowTaskViewButton" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Remove Task View from the Taskbar</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>15</Order>
<CommandLine>reg.exe add "HKCU\SOFTWARE\Microsoft\Windows\CurrentVersion\Explorer\Advanced" /v "TaskbarDa" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Remove Widgets from the Taskbar</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>16</Order>
<CommandLine>reg.exe add "HKCU\SOFTWARE\Microsoft\Windows\CurrentVersion\Explorer\Advanced" /v "TaskbarMn" /t REG_DWORD /d 0 /f</CommandLine>
<Description>Remove Chat from the Taskbar</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>17</Order>
<CommandLine>reg.exe add "HKLM\SOFTWARE\Policies\Microsoft\Windows\WindowsUpdate\AU" /v "NoAutoUpdate" /t REG_DWORD /d 1 /f</CommandLine>
<Description>Turn off Windows Update auto download</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>18</Order>
<CommandLine>netsh advfirewall firewall set rule group="@FirewallAPI.dll,-32752" new enable=Yes</CommandLine>
<Description>Enable Network Discovery</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>19</Order>
<CommandLine>netsh advfirewall firewall set rule group="@FirewallAPI.dll,-28502" new enable=Yes</CommandLine>
<Description>Enable File Sharing</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>20</Order>
<CommandLine>reg.exe add "HKCU\Control Panel\UnsupportedHardwareNotificationCache" /v SV1 /d 0 /t REG_DWORD /f</CommandLine>
<Description>Disable unsupported hardware notifications</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>21</Order>
<CommandLine>reg.exe add "HKCU\Control Panel\UnsupportedHardwareNotificationCache" /v SV2 /d 0 /t REG_DWORD /f</CommandLine>
<Description>Disable unsupported hardware notifications</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>22</Order>
<CommandLine>pnputil -i -a C:\Windows\Drivers\viogpudo\viogpudo.inf</CommandLine>
<Description>Install VirtIO display driver</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>23</Order>
<CommandLine>cmd /C rd /q C:\Windows.old</CommandLine>
<Description>Remove empty Windows.old folder</Description>
</SynchronousCommand>
<SynchronousCommand wcm:action="add">
<Order>24</Order>
<CommandLine>cmd /C if exist "C:\OEM\install.bat" start "Install" "cmd /C C:\OEM\install.bat"</CommandLine>
<Description>Execute custom script from the OEM folder if exists</Description>
</SynchronousCommand>
</FirstLogonCommands>
</component>
</settings>
</unattend>

1
omnitool/omnibox/vm/win11iso/README.md Normal file
View File

@@ -0,0 +1 @@
Add your Win11E setup.iso to this folder

31
omnitool/omnibox/vm/win11setup/firstboot/install.bat Normal file
View File

@@ -0,0 +1,31 @@
@echo off
SET ScriptFolder=\\host.lan\Data
SET LogFile=%ScriptFolder%\firstboot_log.txt
echo Running PowerShell script... > %LogFile%
:: Check for PowerShell availability
where powershell >> %LogFile% 2>&1
if %ERRORLEVEL% neq 0 (
echo PowerShell is not available! >> %LogFile%
echo PowerShell is not available!
exit /b 1
)
:: Add a 30-second delay
echo Waiting for 30 seconds before continuing... >> %LogFile%
timeout /t 30 /nobreak >> %LogFile% 2>&1
:: Run PowerShell script with ExecutionPolicy Bypass and log errors
echo Running setup.ps1... >> %LogFile%
powershell -ExecutionPolicy Bypass -File "%ScriptFolder%\setup.ps1" >> %LogFile% 2>&1
if %ERRORLEVEL% neq 0 (
echo An error occurred. See %LogFile% for details.
) else (
echo PowerShell script has completed successfully.
)
echo PowerShell script has completed.

7
omnitool/omnibox/vm/win11setup/setupscripts/on-logon.ps1 Normal file
View File

@@ -0,0 +1,7 @@
$scriptFolder = "\\host.lan\Data"
$pythonScriptFile = "$scriptFolder\server\main.py"
$pythonServerPort = 5000
# Start the flask computer use server
Write-Host "Running the server on port $pythonServerPort"
python $pythonScriptFile --port $pythonServerPort

BIN
omnitool/omnibox/vm/win11setup/setupscripts/server/cursor.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 3.1 KiB

81
omnitool/omnibox/vm/win11setup/setupscripts/server/main.py Normal file
View File

@@ -0,0 +1,81 @@
import os
import logging
import argparse
import shlex
import subprocess
from flask import Flask, request, jsonify, send_file
import threading
import traceback
import pyautogui
from PIL import Image
from io import BytesIO
parser = argparse.ArgumentParser()
parser.add_argument("--log_file", help="log file path", type=str,
default=os.path.join(os.path.dirname(__file__), "server.log"))
parser.add_argument("--port", help="port", type=int, default=5000)
args = parser.parse_args()
logging.basicConfig(filename=args.log_file,level=logging.DEBUG, filemode='w' )
logger = logging.getLogger('werkzeug')
app = Flask(__name__)
computer_control_lock = threading.Lock()
@app.route('/probe', methods=['GET'])
def probe_endpoint():
return jsonify({"status": "Probe successful", "message": "Service is operational"}), 200
@app.route('/execute', methods=['POST'])
def execute_command():
# Only execute one command at a time
with computer_control_lock:
data = request.json
# The 'command' key in the JSON request should contain the command to be executed.
shell = data.get('shell', False)
command = data.get('command', "" if shell else [])
if isinstance(command, str) and not shell:
command = shlex.split(command)
# Expand user directory
for i, arg in enumerate(command):
if arg.startswith("~/"):
command[i] = os.path.expanduser(arg)
# Execute the command without any safety checks.
try:
result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=shell, text=True, timeout=120)
return jsonify({
'status': 'success',
'output': result.stdout,
'error': result.stderr,
'returncode': result.returncode
})
except Exception as e:
logger.error("\n" + traceback.format_exc() + "\n")
return jsonify({
'status': 'error',
'message': str(e)
}), 500
@app.route('/screenshot', methods=['GET'])
def capture_screen_with_cursor():
cursor_path = os.path.join(os.path.dirname(__file__), "cursor.png")
screenshot = pyautogui.screenshot()
cursor_x, cursor_y = pyautogui.position()
cursor = Image.open(cursor_path)
# make the cursor smaller
cursor = cursor.resize((int(cursor.width / 1.5), int(cursor.height / 1.5)))
screenshot.paste(cursor, (cursor_x, cursor_y), cursor)
# Convert PIL Image to bytes and send
img_io = BytesIO()
screenshot.save(img_io, 'PNG')
img_io.seek(0)
return send_file(img_io, mimetype='image/png')
if __name__ == '__main__':
app.run(debug=True, host="0.0.0.0", port=args.port)

2
omnitool/omnibox/vm/win11setup/setupscripts/server/requirements.txt Normal file
View File

@@ -0,0 +1,2 @@
flask
PyAutoGUI

197
omnitool/omnibox/vm/win11setup/setupscripts/setup-tools.psm1 Normal file
View File

@@ -0,0 +1,197 @@
function Get-Tools {
param(
[string]$toolsConfigJson
)
# Convert the JSON string to a PowerShell object
$toolsList = $toolsConfigJson | ConvertFrom-Json
return $toolsList
}
function Get-ToolDetails {
param(
$toolsList,
[string]$toolName
)
# Check if the program exists in the JSON data
if ($toolsList.PSObject.Properties.Name -contains $toolName) {
# Return the program details as a PowerShell object
return $toolsList.$toolName
} else {
# Handle the case where the program is not found
Write-Host "Program '$toolName' not found in the list."
return $null
}
}
function Invoke-DownloadFileFromAvailableMirrors {
param (
[string[]]$mirrorUrls,
[string]$outfile
)
foreach ($url in $mirrorUrls) {
try {
$result = Invoke-DownloadFile -url $url -outfile $outfile
if ($result -eq $true) {
Write-Host "Downloaded using $url"
return $true
}
} catch {
Write-Host "Error downloading from $url. Please check and update the mirrors."
}
}
Write-Host "Downloading from the provided mirrors failed. Please check and update the mirrors."
return $false
}
function Invoke-DownloadFile {
param (
[string]$url,
[string]$outfile
)
# Makes download faster by disabling progress bar
$ProgressPreference = "SilentlyContinue"
$retryCount = 0
$maxRetries = 3
$sleepSeconds = 2
$maxSleepSeconds = 10
$userAgent = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/58.0.3029.110 Safari/537.36"
# Ensure directory exists
$directory = Split-Path -Path $outfile -Parent
if (-Not (Test-Path -Path $directory)) {
Write-Host "Creating directory $directory..."
New-Item -Path $directory -ItemType Directory -Force | Out-Null
}
while ($retryCount -lt $maxRetries) {
try {
Invoke-RestMethod -Uri $url -OutFile $outfile -Headers @{"User-Agent" = $userAgent}
Write-Host "Download successful, file saved to: $outfile"
break
} catch {
$retryCount++
Write-Host "Attempt $retryCount of $maxRetries failed. Error: $($_.Exception.Message)"
Start-Sleep -Seconds $sleepSeconds
$sleepSeconds = [Math]::Min($sleepSeconds * 2, $maxSleepSeconds) # Exponential backoff with a cap
}
}
if ($retryCount -eq $maxRetries) {
Write-Host "Failed to download the file after $maxRetries attempts."
return $false
}
return $true
}
function Add-ToEnvPath {
param (
[string]$NewPath
)
# Get the current PATH environment variable
$envPath = [Environment]::GetEnvironmentVariable("PATH", "Machine")
# Append the new path to the existing PATH
$newPath = "$envPath;$NewPath"
# Set the updated PATH environment variable
[Environment]::SetEnvironmentVariable("PATH", $newPath, "Machine")
# Fetch updates from the shell
$env:PATH += ";${newPath}"
}
function Register-LogonTask {
param(
[parameter(Mandatory = $true, ValueFromPipelineByPropertyName = $true, HelpMessage = "Name of the scheduled task")]
[string]
$TaskName,
[parameter(Mandatory = $true, ValueFromPipelineByPropertyName = $true, HelpMessage = "Path to the .py script")]
[string]
$ScriptPath,
[parameter(Mandatory = $false, ValueFromPipelineByPropertyName = $true, HelpMessage = "Arguments to the .py script")]
[string]
$Arguments = "",
[parameter(Mandatory = $false, ValueFromPipelineByPropertyName = $true, HelpMessage = "Local Account username")]
[string]
$LocalUser,
[parameter(Mandatory = $false, ValueFromPipelineByPropertyName = $true, HelpMessage = "Local Account password")]
[string]
$LocalPassword,
[parameter(Mandatory = $false, ValueFromPipelineByPropertyName = $true, HelpMessage = "Whether to execute the command as SYSTEM")]
[switch]
$AsSystem = $false,
[parameter(Mandatory = $false, ValueFromPipelineByPropertyName = $true, HelpMessage = "logging file")]
[string]
$LogFilePath
)
$scriptDirectory = Split-Path $ScriptPath
$taskActionArgument = "-ExecutionPolicy Bypass -windowstyle hidden -Command `"try { . '$ScriptPath' $Arguments } catch { Write `$_.Exception.Message | Out-File $($TaskName)_Log.txt } finally { } `""
$taskAction = New-ScheduledTaskAction -Execute "$PSHome\powershell.exe" -Argument $taskActionArgument -WorkingDirectory $scriptDirectory
$params = @{
Force = $True
Action = $taskAction
RunLevel = "Highest"
TaskName = $TaskName
}
$taskTrigger = New-ScheduledTaskTrigger -AtLogOn
$params.Add("Trigger", $taskTrigger)
if ($AsSystem) {
$params.Add("User", "NT AUTHORITY\SYSTEM")
}
else {
$params.Add("User", $LocalUser)
if ($LocalPassword) {
$params.Add("Password", $LocalPassword)
}
}
Write-Host "Registering scheduled task '$TaskName' to run 'powershell.exe $taskActionArgument'..."
Register-ScheduledTask @params
}
# Function to attempt pip install and handle failures
function Install-PythonPackages {
param (
[string]$Package = "",
[string]$Arguments = "",
[string]$RequirementsPath = ""
)
$RetryCount = 3
$currentAttempt = 0
while ($currentAttempt -lt $RetryCount) {
if (-not [string]::IsNullOrWhiteSpace($RequirementsPath)) {
& python -m pip install --no-cache-dir -r $RequirementsPath $Arguments
} else {
& python -m pip install --no-cache-dir $Package $Arguments
}
if ($LASTEXITCODE -eq 0) {
Write-Host "Installation successful."
return
} else {
Write-Host "Attempt $($currentAttempt + 1) failed. Retrying..."
Start-Sleep -Seconds 10
$currentAttempt++
}
}
Write-Error "Failed to install after $RetryCount attempts."
exit
}

392
omnitool/omnibox/vm/win11setup/setupscripts/setup.ps1 Normal file
View File

@@ -0,0 +1,392 @@
$ErrorActionPreference = "Continue" # until downloading from mirrors is more stable
# Section - General Setup
$scriptFolder = "\\host.lan\Data"
$toolsFolder = "C:\Users\$env:USERNAME\Tools"
# Load the shared setup-tools module
Import-Module (Join-Path $scriptFolder -ChildPath "setup-tools.psm1")
# Check if profile exists
if (-not (Test-Path $PROFILE)) {
New-Item -ItemType File -Path $PROFILE -Force
}
# Create a folder where we store all the standalone executables
if (-not (Test-Path $toolsFolder)) {
New-Item -ItemType Directory -Path $toolsFolder -Force
$envPath = [Environment]::GetEnvironmentVariable("PATH", "Machine")
$newPath = "$envPath;$toolsFolder"
[Environment]::SetEnvironmentVariable("PATH", $newPath, "Machine")
}
# Section - Tools Installation
# Set TLS version to 1.2 or higher
[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12 -bor [Net.SecurityProtocolType]::Tls13
# Load the tools config json listing mirrors and aliases used for installing tools
$toolsConfigJsonPath = Join-Path $scriptFolder -ChildPath "tools_config.json"
$toolsConfigJson = Get-Content -Path $toolsConfigJsonPath -Raw
$toolsList = Get-Tools -toolsConfigJson $toolsConfigJson
## - Python
$pythonToolName = "Python"
$userPythonPath = "$env:LOCALAPPDATA\Programs\Python"
$pythonDetails = Get-ToolDetails -toolsList $toolsList -toolName $pythonToolName
$pythonAlias = $pythonDetails.alias
# Check for Python installation
$pythonExecutablePath = Get-ChildItem -Path $userPythonPath -Filter python.exe -Recurse -ErrorAction SilentlyContinue | Select-Object -First 1 -ExpandProperty FullName
# Force to install Python 3.10 as the pre-installed version on Windows may not work sometimes
Write-Host "Downloading Python $pythonVersion..."
$pythonInstallerFilePath = "$env:TEMP\python_installer.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $pythonDetails.mirrors -outfile $pythonInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download Python. Please try again later or install manually."
} else {
Write-Host "Installing Python for current user..."
Start-Process -FilePath $pythonInstallerFilePath -Args "/quiet InstallAllUsers=0 PrependPath=0" -NoNewWindow -Wait
$pythonExecutablePath = "$userPythonPath\Python310\python.exe"
$setAliasExpression = "Set-Alias -Name $pythonAlias -Value `"$pythonExecutablePath`""
Add-Content -Path $PROFILE -Value $setAliasExpression
Invoke-Expression $setAliasExpression
}
## - Git
$gitToolName = "git"
$gitToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $gitToolName
# Check for Git installation
try {
git --version | Out-Null
Write-Host "Git is already installed."
} catch {
Write-Host "Git is not installed. Downloading and installing Git..."
$gitInstallerFilePath = "$env:TEMP\git_installer.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $gitToolDetails.mirrors -outfile $gitInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download Git. Please try again later or install manually."
} else {
Start-Process -FilePath $gitInstallerFilePath -Args "/VERYSILENT /NORESTART /NOCANCEL /SP-" -Wait
Add-ToEnvPath -NewPath "C:\Program Files\Git\bin"
Write-Host "Git has been installed."
}
}
# - 7zip
$7ZipToolName = "7zip"
$7ZipToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $7ZipToolName
Write-Host "$7ZipToolDetails"
if (Get-Command 7z -ErrorAction SilentlyContinue) {
Write-Host "7-Zip is already installed."
}
else {
Write-Host "Installing 7-Zip..."
$7ZipInstallerFilePath = "$env:TEMP\7_zip.exe"
Write-Host "$($7ZipToolDetails.mirrors)"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $7ZipToolDetails.mirrors -outfile $7ZipInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download 7-Zip. Please try again later or install manually."
} else {
Start-Process -FilePath $7ZipInstallerFilePath -Args "/S" -Verb RunAs -Wait
Remove-Item $7ZipInstallerFilePath
# add 7z to PATH
Add-ToEnvPath -NewPath "${env:ProgramFiles}\7-Zip"
}
}
# - ffpmeg
$ffpmegToolName = "ffmpeg"
$ffpmegToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $ffpmegToolName
if (Get-Command ffmpeg -ErrorAction SilentlyContinue) {
Write-Host "ffmpeg is already installed."
} else {
Write-Host "ffmpeg is not installed. Installing it."
$ffpmegInstallerFilePath = "C:\ffmpeg.7z"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $ffpmegToolDetails.mirrors -outfile $ffpmegInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download ffmpeg. Please try again later or install manually."
} else {
Write-Host "Extracting $ffpmegInstallerFilePath..."
7z x -y -o"C:\" "C:\ffmpeg.7z"
$ffmpegFolder = Get-ChildItem -Path "C:\" -Filter "ffmpeg-*" -Directory
$ffmpegFolder = -join ("C:\", $ffmpegFolder)
#remove ffmpeg folder if exists
if (Test-Path "C:\ffmpeg") {
Remove-Item -Path "C:\ffmpeg" -Recurse -Force
}
Rename-Item -Path "$ffmpegFolder" -NewName "ffmpeg"
Write-Host "Adding ffmpeg to PATH..."
Add-ToEnvPath -NewPath "C:\ffmpeg\bin"
Write-Host "ffmpeg is installed"
}
}
# Disable Edge Auto Updates
Stop-Process -Name "MicrosoftEdgeUpdate" -Force -ErrorAction SilentlyContinue
$edgeUpdatePath = "${env:ProgramFiles(x86)}\Microsoft\EdgeUpdate"
Remove-Item -Path $edgeUpdatePath -Recurse -Force -ErrorAction SilentlyContinue
Write-Host "Edge Update processes terminated and directory removed."
# - Google Chrome
$chromeToolName = "Google Chrome"
$chromeToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $chromeToolName
$chromeExePath = "C:\Program Files\Google\Chrome\Application\chrome.exe"
$chromeAlias = $chromeToolDetails.alias
# Check if Google Chrome is already installed by its alias
if (Get-Command $chromeAlias -ErrorAction SilentlyContinue) {
Write-Host "Google Chrome is already installed."
} else {
# Download the installer to the Temp directory
$chromeInstallerFilePath = "$env:TEMP\chrome_installer.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $chromeToolDetails.mirrors -outfile $chromeInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download Google Chrome. Please try again later or install manually."
} else {
# Execute the installer silently with elevated permissions
Start-Process -FilePath $chromeInstallerFilePath -ArgumentList "/silent", "/install" -Verb RunAs -Wait
# Remove the installer file after installation
Remove-Item -Path $chromeInstallerFilePath
# Set alias
$setAliasExpression = "Set-Alias -Name $chromeAlias -Value `"$chromeExePath`""
Add-Content -Path $PROFILE -Value $setAliasExpression
Invoke-Expression $setAliasExpression
# Add Chrome to the system PATH environment variable
Add-ToEnvPath -NewPath "${env:ProgramFiles}\Google\Chrome\Application"
# Disable Google Chrome Auto Updates
$chromeRegPath = "HKLM:\SOFTWARE\Policies\Google\Update"
if (-not (Test-Path $chromeRegPath)) {
New-Item -Path $chromeRegPath -Force
}
Set-ItemProperty -Path $chromeRegPath -Name "AutoUpdateCheckPeriodMinutes" -Value 0
Set-ItemProperty -Path $chromeRegPath -Name "UpdateDefault" -Value 0
}
}
# - LibreOffice
$libreOfficeToolName = "LibreOffice"
$libreOfficeToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $libreOfficeToolName
# Check for LibreOffice installation
$installedVersion = (Get-WmiObject -Query "SELECT * FROM Win32_Product WHERE Name like 'LibreOffice%'").Version
if (-not [string]::IsNullOrWhiteSpace($installedVersion)) {
Write-Host "LibreOffice $version is already installed."
} else {
Write-Host "LibreOffice is not installed. Downloading and installing LibreOffice..."
$libreOfficeInstallerFilePath = "$env:TEMP\libreOffice_installer.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $libreOfficeToolDetails.mirrors -outfile $libreOfficeInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download LibreOffice. Please try again later or install manually."
} else {
Start-Process "msiexec.exe" -ArgumentList "/i `"$libreOfficeInstallerFilePath`" /quiet" -Wait -NoNewWindow
Write-Host "LibreOffice has been installed."
# Add LibreOffice to the system PATH environment variable
Add-ToEnvPath -NewPath "C:\Program Files\LibreOffice\program"
}
}
# - VLC
$vlcToolName = "VLC"
$vlcToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $vlcToolName
$vlcAlias = $vlcToolDetails.alias
$vlcExecutableFilePath = "C:\Program Files\VideoLAN\VLC\vlc.exe"
# Check if VLC is already installed by checking the VLC command
if (Test-Path $vlcExecutableFilePath) {
Write-Host "VLC is already installed."
} else {
# Download the installer to the Temp directory
$vlcInstallerFilePath = "$env:TEMP\vlc_installer.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $vlcToolDetails.mirrors -outfile $vlcInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download VLC. Please try again later or install manually."
} else {
# Execute the installer silently with elevated permissions
Start-Process -FilePath $vlcInstallerFilePath -ArgumentList "/S" -Verb RunAs -Wait
# Remove the installer file after installation
Remove-Item -Path $vlcInstallerFilePath
# Set alias
$setAliasExpression = "Set-Alias -Name $vlcAlias -Value `"$vlcExecutableFilePath`""
Add-Content -Path $PROFILE -Value $setAliasExpression
Invoke-Expression $setAliasExpression
# Add VLC to the system PATH environment variable
Add-ToEnvPath -NewPath "C:\Program Files\VideoLAN\VLC"
}
}
# - GIMP
$gimpToolName = "GIMP"
$gimpToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $gimpToolName
$gimpAlias = $gimpToolDetails.alias
$gimpExecutablePath = "C:\Program Files\GIMP 2\bin\gimp-2.10.exe"
# Check if GIMP is already installed by checking the GIMP executable path
if (Test-Path $gimpExecutablePath) {
Write-Host "GIMP is already installed."
} else {
# Download the installer to the Temp directory
$gimpInstallerFilePath = "$env:TEMP\gimp_installer.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $gimpToolDetails.mirrors -outfile $gimpInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download GIMP. Please try again later or install manually."
} else {
# Execute the installer silently with elevated permissions
Start-Process -FilePath $gimpInstallerFilePath -ArgumentList "/VERYSILENT /ALLUSERS" -Verb RunAs -Wait
# Remove the installer file after installation
Remove-Item -Path $gimpInstallerFilePath
# Set alias
$setAliasExpression = "Set-Alias -Name $gimpAlias -Value `"$gimpExecutablePath`""
Add-Content -Path $PROFILE -Value $setAliasExpression
Invoke-Expression $setAliasExpression
# Add GIMP to the system PATH environment variable
Add-ToEnvPath -NewPath "C:\Program Files\GIMP 2\bin"
}
}
# - VS Code
$vsCodeToolName = "VS Code"
$vsCodeToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $vsCodeToolName
$vsCodeAlias = $gimpToolDetails.alias
$vsCodeExecutablePath = "C:\Users\$env:USERNAME\AppData\Local\Programs\Microsoft VS Code\Code.exe"
# Check if VS Code is already installed by checking the VS Code executable path
if (Test-Path $vsCodeExecutablePath) {
Write-Host "VS Code is already installed."
} else {
# Download the installer to the Temp directory
$vsCodeInstallerFilePath = "$env:TEMP\VSCodeSetup.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $vsCodeToolDetails.mirrors -outfile $vsCodeInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download VS Code. Please try again later or install manually."
} else {
# Execute the installer silently with elevated permissions
Start-Process -FilePath $vsCodeInstallerFilePath -ArgumentList "/VERYSILENT", "/mergetasks=!runcode" -Verb RunAs -Wait
# Remove the installer file after installation
Remove-Item -Path $vsCodeInstallerFilePath
# Set alias
$setAliasExpression = "Set-Alias -Name $vsCodeAlias -Value `"$vsCodeExecutablePath`""
Add-Content -Path $PROFILE -Value $setAliasExpression
Invoke-Expression $setAliasExpression
# Add VS Code to the system PATH environment variable
Add-ToEnvPath -NewPath "C:\Users\$env:USERNAME\AppData\Local\Programs\Microsoft VS Code\bin"
# Disable Visual Studio Code Auto Updates
$vsCodeSettingsPath = "${env:APPDATA}\Code\User\settings.json"
if (-not (Test-Path $vsCodeSettingsPath)) {
# Create the directory if it doesn't exist
$dirPath = Split-Path -Path $vsCodeSettingsPath -Parent
if (-not (Test-Path $dirPath)) {
New-Item -ItemType Directory -Path $dirPath -Force
}
# Initialize an empty hashtable to act as the JSON object
$settingsObj = @{}
$settingsObj["update.mode"] = "none" # Set update mode to none
$settingsObj | ConvertTo-Json | Set-Content $vsCodeSettingsPath
} else {
# If the file exists, modify it
$settingsObj = Get-Content $vsCodeSettingsPath | ConvertFrom-Json
$settingsObj["update.mode"] = "none"
$settingsObj | ConvertTo-Json | Set-Content $vsCodeSettingsPath
}
}
}
# - Thunderbird
$thunderbirdToolName = "Thunderbird"
$thunderbirdToolDetails = Get-ToolDetails -toolsList $toolsList -toolName $thunderbirdToolName
$thunderbirdAlias = $thunderbirdToolDetails.alias
$thunderbirdExecutablePath = "C:\Program Files\Mozilla Thunderbird\thunderbird.exe"
# Check if Thunderbird is already installed by checking the Thunderbird executable path
if (Test-Path $thunderbirdExecutablePath) {
Write-Host "Thunderbird is already installed."
} else {
# Download the installer to the Temp directory
$thunderbirdInstallerFilePath = "$env:TEMP\ThunderbirdSetup.exe"
$downloadResult = Invoke-DownloadFileFromAvailableMirrors -mirrorUrls $thunderbirdToolDetails.mirrors -outfile $thunderbirdInstallerFilePath
if (-not $downloadResult) {
Write-Host "Failed to download Thunderbird. Please try again later or install manually."
} else {
# Execute the installer silently with elevated permissions
Start-Process -FilePath $thunderbirdInstallerFilePath -ArgumentList "/S" -Verb RunAs -Wait
# Remove the installer file after installation
Remove-Item -Path $thunderbirdInstallerFilePath
# Set alias
$setAliasExpression = "Set-Alias -Name $thunderbirdAlias -Value `"$thunderbirdExecutablePath`""
Add-Content -Path $PROFILE -Value $setAliasExpression
Invoke-Expression $setAliasExpression
# Add Thunderbird to the system PATH environment variable
Add-ToEnvPath -NewPath "C:\Program Files\Mozilla Thunderbird"
}
}
# - Server Setup
$pythonServerPort = 5000
$onLogonTaskName = "Server_OnLogon"
$requirementsFile = "$scriptFolder\server\requirements.txt"
# Ensure pip is updated to the latest version
Install-PythonPackages -Package "pip" -Arguments "--upgrade"
Install-PythonPackages -Package "wheel"
Install-PythonPackages -Package "pywinauto"
# Install Python packages from requirements.txt using Python's pip module
if (Test-Path $requirementsFile) {
Write-Host "Installing required Python packages using pip from requirements file..."
Install-PythonPackages -RequirementsPath $requirementsFile
} else {
Write-Error "Requirements file not found: $requirementsFile"
exit
}
# Add a firewall rule to allow incoming connections on the specified port for the Python executable
$pythonServerRuleName = "PythonHTTPServer-$pythonServerPort"
if (-not (Get-NetFirewallRule -Name $pythonServerRuleName -ErrorAction SilentlyContinue)) {
New-NetFirewallRule -DisplayName $pythonServerRuleName -Direction Inbound -Program $pythonExecutablePath -Protocol TCP -LocalPort $pythonServerPort -Action Allow -Profile Any
Write-Host "Firewall rule added to allow traffic on port $pythonServerPort for Python"
} else {
Write-Host "Firewall rule already exists. $pythonServerRuleName "
}
$onLogonScriptPath = "$scriptFolder\on-logon.ps1"
# Check if the scheduled task exists before unregistering it
if (Get-ScheduledTask -TaskName $onLogonTaskName -ErrorAction SilentlyContinue) {
Write-Host "Scheduled task $onLogonTaskName already exists."
} else {
Write-Host "Registering new task $onLogonTaskName..."
Register-LogonTask -TaskName $onLogonTaskName -ScriptPath $onLogonScriptPath -LocalUser "Docker"
}
Start-Sleep -Seconds 10
Start-ScheduledTask -TaskName $onLogonTaskName

71
omnitool/omnibox/vm/win11setup/setupscripts/tools_config.json Normal file
View File

@@ -0,0 +1,71 @@
{
"Python": {
"mirrors": [
"https://www.python.org/ftp/python/3.10.0/python-3.10.0-amd64.exe"
],
"alias": "python"
},
"git": {
"mirrors": [
"https://github.com/git-for-windows/git/releases/download/v2.37.1.windows.1/Git-2.37.1-64-bit.exe"
]
},
"7zip": {
"mirrors": [
"https://www.7-zip.org/a/7z2407-x64.exe"
]
},
"ffmpeg": {
"mirrors": [
"https://www.gyan.dev/ffmpeg/builds/ffmpeg-release-essentials.7z"
]
},
"Google Chrome": {
"mirrors": [
"https://dl.google.com/chrome/install/latest/chrome_installer.exe"
],
"alias": "google-chrome"
},
"LibreOffice": {
"mirrors": [
"https://mirror.raiolanetworks.com/tdf/libreoffice/stable/24.8.4/win/x86_64/LibreOffice_24.8.4_Win_x86-64.msi",
"https://mirrors.iu13.net/tdf/libreoffice/stable/24.8.4/win/x86_64/LibreOffice_24.8.4_Win_x86-64.msi",
"https://download.documentfoundation.org/libreoffice/stable/24.8.4/win/x86_64/LibreOffice_24.8.4_Win_x86-64.msi"
]
},
"VLC": {
"mirrors": [
"https://ftp.free.org/mirrors/videolan/vlc/3.0.21/win64/vlc-3.0.21-win64.exe",
"https://mirror.fcix.net/videolan-ftp/vlc/3.0.21/win64/vlc-3.0.21-win64.exe",
"https://mirror.raiolanetworks.com/videolan/vlc/3.0.21/win64/vlc-3.0.21-win64.exe"
],
"alias": "vlc"
},
"GIMP": {
"mirrors": [
"https://www-ftp.lip6.fr/pub/gimp/gimp/v2.10/windows/gimp-2.10.38-setup.exe",
"https://download.gimp.org/gimp/v2.10/windows/gimp-2.10.38-setup.exe",
"https://www-ftp.lip6.fr/pub/gimp/gimp/v2.10/windows/gimp-2.10.0-setup.exe"
],
"alias": "gimp"
},
"VS Code": {
"mirrors": [
"https://update.code.visualstudio.com/latest/win32-x64-user/stable"
],
"alias": "code"
},
"Thunderbird": {
"mirrors": [
"https://download-installer.cdn.mozilla.net/pub/thunderbird/releases/115.12.1/win64/en-US/Thunderbird%20Setup%20115.12.1.exe",
"https://archive.mozilla.org/pub/thunderbird/releases/115.12.1/win64/en-US/Thunderbird%20Setup%20115.12.1.exe"
],
"alias": "thunderbird"
},
"Caddy Proxy": {
"mirrors": [
"https://caddyserver.com/api/download?os=windows&arch=amd64"
],
"alias": "caddy"
}
}

51
omnitool/omniparserserver/omniparserserver.py Normal file
View File

@@ -0,0 +1,51 @@
'''
python -m omniparserserver --som_model_path ../../weights/icon_detect_v1_5/model_v1_5.pt --caption_model_name florence2 --caption_model_path ../../weights/icon_caption_florence --device cuda --BOX_TRESHOLD 0.05
'''
import sys
import os
import time
from fastapi import FastAPI
from pydantic import BaseModel
import argparse
import uvicorn
root_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
sys.path.append(root_dir)
from util.omniparser import Omniparser
def parse_arguments():
parser = argparse.ArgumentParser(description='Omniparser API')
parser.add_argument('--som_model_path', type=str, default='../../weights/icon_detect_v1_5/model_v1_5.pt', help='Path to the som model')
parser.add_argument('--caption_model_name', type=str, default='florence2', help='Name of the caption model')
parser.add_argument('--caption_model_path', type=str, default='../../weights/icon_caption_florence', help='Path to the caption model')
parser.add_argument('--device', type=str, default='cpu', help='Device to run the model')
parser.add_argument('--BOX_TRESHOLD', type=float, default=0.05, help='Threshold for box detection')
parser.add_argument('--host', type=str, default='0.0.0.0', help='Host for the API')
parser.add_argument('--port', type=int, default=8000, help='Port for the API')
args = parser.parse_args()
return args
args = parse_arguments()
config = vars(args)
app = FastAPI()
omniparser = Omniparser(config)
class ParseRequest(BaseModel):
base64_image: str
@app.post("/parse/")
async def parse(parse_request: ParseRequest):
print('start parsing...')
start = time.time()
dino_labled_img, parsed_content_list = omniparser.parse(parse_request.base64_image)
latency = time.time() - start
print('time:', latency)
return {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, 'latency': latency}
@app.get("/probe/")
async def root():
return {"message": "Omniparser API ready"}
if __name__ == "__main__":
uvicorn.run("omniparserserver:app", host=args.host, port=args.port, reload=True)

95
omnitool/readme.md Normal file
View File

@@ -0,0 +1,95 @@
<img src="../imgs/header_bar.png" alt="OmniTool Header" width="100%">
# OmniTool
Control a Windows 11 VM with OmniParser + your vision model of choice.
## Highlights:
1. **OmniParser V2** is 60% faster than V1 and now understands a wide variety of OS, app and inside app icons!
2. **OmniBox** uses 50% less disk space than other Windows VMs for agent testing, whilst providing the same computer use API
3. **OmniTool** supports out of the box the following vision models - OpenAI (4o/o1/o3-mini), DeepSeek (R1), Qwen (2.5VL) or Anthropic Computer Use
## Overview
There are three components:
<table style="border-collapse: collapse; border: none;">
<tr>
<td style="border: none;"><img src="../imgs/omniparsericon.png" width="50"></td>
<td style="border: none;"><strong>omniparserserver</strong></td>
<td style="border: none;">FastAPI server running OmniParser V2.</td>
</tr>
<tr>
<td style="border: none;"><img src="../imgs/omniboxicon.png" width="50"></td>
<td style="border: none;"><strong>omnibox</strong></td>
<td style="border: none;">A Windows 11 VM running in a Docker container.</td>
</tr>
<tr>
<td style="border: none;"><img src="../imgs/gradioicon.png" width="50"></td>
<td style="border: none;"><strong>gradio</strong></td>
<td style="border: none;">UI to provide commands and watch reasoning + execution on OmniBox</td>
</tr>
</table>
## Notes:
1. Though **OmniParser V2** can run on a CPU, we have separated this out if you want to run it fast on a GPU machine
2. The **OmniBox** Windows 11 VM docker is dependent on KVM so can only run quickly on Windows and Linux. This can run on a CPU machine (doesn't need GPU).
3. The Gradio UI can also run on a CPU machine. We suggest running **omnibox** and **gradio** on the same CPU machine and **omniparserserver** on a GPU server.
## Setup
1. **omniparserserver**:
a. If you already have a conda environment for OmniParser, you can use that. Else follow the following steps to create one
b. Ensure conda is installed with `conda --version` or install from the [Anaconda website](https://www.anaconda.com/download/success)
c. Navigate to the root of the repo with `cd OmniParser`
d. Create a conda python environment with `conda create -n "omni" python==3.12`
e. Set the python environment to be used with `conda activate omni`
f. Install the dependencies with `pip install -r requirements.txt`
g. Continue from here if you already had the conda environment.
h. Ensure you have the weights downloaded in weights folder. If not download them with:
`for folder in icon_caption_florence icon_detect icon_detect_v1_5; do huggingface-cli download microsoft/OmniParser --local-dir weights/ --repo-type model --include "$folder/*"; done`
h. Navigate to the server directory with `cd OmniParser/omnitool/omniparserserver`
i. Start the server with `python -m omniparserserver`
2. **omnibox**:
a. Install Docker Desktop
b. Visit [Microsoft Evaluation Center](https://info.microsoft.com/ww-landing-windows-11-enterprise.html), accept the Terms of Service, and download a **Windows 11 Enterprise Evaluation (90-day trial, English, United States)** ISO file [~6GB]. Rename the file to `custom.iso` and copy it to the directory `OmniParser/omnitool/omnibox/vm/win11iso`
c. Navigate to vm management script directory with`cd OmniParser/omnitool/omnibox/scripts`
d. Build the docker container [400MB] and install the ISO to a storage folder [20GB] with `./manage_vm.sh create`
e. After creating the first time it will store a save of the VM state in `vm/win11storage`. You can then manage the VM with `./manage_vm.sh start` and `./manage_vm.sh stop`. To delete the VM, use `./manage_vm.sh delete` and delete the `OmniParser/omnitool/omnibox/vm/win11storage` directory.
3. **gradio**:
a. Navigate to the gradio directory with `cd OmniParser/omnitool/gradio`
b. Ensure you have activated the conda python environment with `conda activate omni`
c. Start the server with `python app.py --windows_host_url localhost:8006 --omniparser_server_url localhost:8000`
d. Open the URL in the terminal output, set your API Key and start playing with the AI agent!
## Risks and Mitigations
To align with the Microsoft AI principles and Responsible AI practices, we conduct risk mitigation by training the icon caption model with Responsible AI data, which helps the model avoid inferring sensitive attributes (e.g.race, religion etc.) of the individuals which happen to be in icon images as much as possible. At the same time, we encourage user to apply OmniParser only for screenshot that does not contain harmful/violent content. For the OmniTool, we conduct threat model analysis using Microsoft Threat Modeling Tool overview - Azure | Microsoft Learn. We provide a sandbox docker container, and provide safety guidance and examples in our GitHub Repository. We advise human to stay in the loop in order to minimize risk.
## Acknowledgment
Kudos to the amazing resources that are indispensable in the development of our code: [Claude Computer Use](https://github.com/anthropics/anthropic-quickstarts/blob/main/computer-use-demo/README.md), [OS World](https://github.com/xlang-ai/OSWorld), [Windows Agent Arena](https://github.com/microsoft/WindowsAgentArena), and [computer_use_ootb](https://github.com/showlab/computer_use_ootb).
We are grateful for helpful suggestions and feedbacks provided by Francesco Bonacci, Jianwei Yang, Dillon DuPont, Yue Wu, Anh Nguyen.

5
requirements.txt
View File

@@ -4,7 +4,7 @@ torchvision
supervision==0.18.0
openai==1.3.5
transformers
ultralytics==8.1.24
ultralytics==8.3.70
azure-identity
numpy
opencv-python
@@ -28,4 +28,5 @@ boto3>=1.28.57
google-auth<3,>=2
screeninfo
uiautomation
dashscope
dashscope
groq

BIN
util/__pycache__/__init__.cpython-312.pyc
View File

Binary file not shown.

BIN
util/__pycache__/__init__.cpython-39.pyc
View File

Binary file not shown.

BIN
util/__pycache__/action_matching.cpython-39.pyc
View File

Binary file not shown.

BIN
util/__pycache__/box_annotator.cpython-312.pyc
View File

Binary file not shown.

BIN
util/__pycache__/box_annotator.cpython-39.pyc
View File

Binary file not shown.

425
util/action_matching.py
View File

@@ -1,425 +0,0 @@
'''
Adapted from https://github.com/google-research/google-research/tree/master/android_in_the_wild
'''
import jax
import jax.numpy as jnp
import numpy as np
# import action_type as action_type_lib
import enum
class ActionType(enum.IntEnum):
# Placeholders for unused enum values
UNUSED_0 = 0
UNUSED_1 = 1
UNUSED_2 = 2
UNUSED_8 = 8
UNUSED_9 = 9
########### Agent actions ###########
# A type action that sends text to the emulator. Note that this simply sends
# text and does not perform any clicks for element focus or enter presses for
# submitting text.
TYPE = 3
# The dual point action used to represent all gestures.
DUAL_POINT = 4
# These actions differentiate pressing the home and back button from touches.
# They represent explicit presses of back and home performed using ADB.
PRESS_BACK = 5
PRESS_HOME = 6
# An action representing that ADB command for hitting enter was performed.
PRESS_ENTER = 7
########### Episode status actions ###########
# An action used to indicate the desired task has been completed and resets
# the environment. This action should also be used in the case that the task
# has already been completed and there is nothing to do.
# e.g. The task is to turn on the Wi-Fi when it is already on
STATUS_TASK_COMPLETE = 10
# An action used to indicate that desired task is impossible to complete and
# resets the environment. This can be a result of many different things
# including UI changes, Android version differences, etc.
STATUS_TASK_IMPOSSIBLE = 11
_TAP_DISTANCE_THRESHOLD = 0.14 # Fraction of the screen
ANNOTATION_WIDTH_AUGMENT_FRACTION = 1.4
ANNOTATION_HEIGHT_AUGMENT_FRACTION = 1.4
# Interval determining if an action is a tap or a swipe.
_SWIPE_DISTANCE_THRESHOLD = 0.04
def _yx_in_bounding_boxes(
yx, bounding_boxes
):
"""Check if the (y,x) point is contained in each bounding box.
Args:
yx: The (y, x) coordinate in pixels of the point.
bounding_boxes: A 2D int array of shape (num_bboxes, 4), where each row
represents a bounding box: (y_top_left, x_top_left, box_height,
box_width). Note: containment is inclusive of the bounding box edges.
Returns:
is_inside: A 1D bool array where each element specifies if the point is
contained within the respective box.
"""
y, x = yx
# `bounding_boxes` has shape (n_elements, 4); we extract each array along the
# last axis into shape (n_elements, 1), then squeeze unneeded dimension.
top, left, height, width = [
jnp.squeeze(v, axis=-1) for v in jnp.split(bounding_boxes, 4, axis=-1)
]
# The y-axis is inverted for AndroidEnv, so bottom = top + height.
bottom, right = top + height, left + width
return jnp.logical_and(y >= top, y <= bottom) & jnp.logical_and(
x >= left, x <= right)
def _resize_annotation_bounding_boxes(
annotation_positions, annotation_width_augment_fraction,
annotation_height_augment_fraction):
"""Resize the bounding boxes by the given fractions.
Args:
annotation_positions: Array of shape (N, 4), where each row represents the
(y, x, height, width) of the bounding boxes.
annotation_width_augment_fraction: The fraction to augment the box widths,
E.g., 1.4 == 240% total increase.
annotation_height_augment_fraction: Same as described for width, but for box
height.
Returns:
Resized bounding box.
"""
height_change = (
annotation_height_augment_fraction * annotation_positions[:, 2])
width_change = (
annotation_width_augment_fraction * annotation_positions[:, 3])
# Limit bounding box positions to the screen.
resized_annotations = jnp.stack([
jnp.maximum(0, annotation_positions[:, 0] - (height_change / 2)),
jnp.maximum(0, annotation_positions[:, 1] - (width_change / 2)),
jnp.minimum(1, annotation_positions[:, 2] + height_change),
jnp.minimum(1, annotation_positions[:, 3] + width_change),
],
axis=1)
return resized_annotations
def is_tap_action(normalized_start_yx,
normalized_end_yx):
distance = jnp.linalg.norm(
jnp.array(normalized_start_yx) - jnp.array(normalized_end_yx))
return distance <= _SWIPE_DISTANCE_THRESHOLD
def _is_non_dual_point_action(action_type):
return jnp.not_equal(action_type, ActionType.DUAL_POINT)
def _check_tap_actions_match(
tap_1_yx,
tap_2_yx,
annotation_positions,
matching_tap_distance_threshold_screen_percentage,
annotation_width_augment_fraction,
annotation_height_augment_fraction,
):
"""Determines if two tap actions are the same."""
resized_annotation_positions = _resize_annotation_bounding_boxes(
annotation_positions,
annotation_width_augment_fraction,
annotation_height_augment_fraction,
)
# Check if the ground truth tap action falls in an annotation's bounding box.
tap1_in_box = _yx_in_bounding_boxes(tap_1_yx, resized_annotation_positions)
tap2_in_box = _yx_in_bounding_boxes(tap_2_yx, resized_annotation_positions)
both_in_box = jnp.max(tap1_in_box & tap2_in_box)
# If the ground-truth tap action falls outside any of the annotation
# bounding boxes or one of the actions is inside a bounding box and the other
# is outside bounding box or vice versa, compare the points using Euclidean
# distance.
within_threshold = (
jnp.linalg.norm(jnp.array(tap_1_yx) - jnp.array(tap_2_yx))
<= matching_tap_distance_threshold_screen_percentage
)
return jnp.logical_or(both_in_box, within_threshold)
def _check_drag_actions_match(
drag_1_touch_yx,
drag_1_lift_yx,
drag_2_touch_yx,
drag_2_lift_yx,
):
"""Determines if two drag actions are the same."""
# Store drag deltas (the change in the y and x coordinates from touch to
# lift), magnitudes, and the index of the main axis, which is the axis with
# the greatest change in coordinate value (e.g. a drag starting at (0, 0) and
# ending at (0.3, 0.5) has a main axis index of 1).
drag_1_deltas = drag_1_lift_yx - drag_1_touch_yx
drag_1_magnitudes = jnp.abs(drag_1_deltas)
drag_1_main_axis = np.argmax(drag_1_magnitudes)
drag_2_deltas = drag_2_lift_yx - drag_2_touch_yx
drag_2_magnitudes = jnp.abs(drag_2_deltas)
drag_2_main_axis = np.argmax(drag_2_magnitudes)
return jnp.equal(drag_1_main_axis, drag_2_main_axis)
def check_actions_match(
action_1_touch_yx,
action_1_lift_yx,
action_1_action_type,
action_2_touch_yx,
action_2_lift_yx,
action_2_action_type,
annotation_positions,
tap_distance_threshold = _TAP_DISTANCE_THRESHOLD,
annotation_width_augment_fraction = ANNOTATION_WIDTH_AUGMENT_FRACTION,
annotation_height_augment_fraction = ANNOTATION_HEIGHT_AUGMENT_FRACTION,
):
"""Determines if two actions are considered to be the same.
Two actions being "the same" is defined here as two actions that would result
in a similar screen state.
Args:
action_1_touch_yx: The (y, x) coordinates of the first action's touch.
action_1_lift_yx: The (y, x) coordinates of the first action's lift.
action_1_action_type: The action type of the first action.
action_2_touch_yx: The (y, x) coordinates of the second action's touch.
action_2_lift_yx: The (y, x) coordinates of the second action's lift.
action_2_action_type: The action type of the second action.
annotation_positions: The positions of the UI annotations for the screen. It
is A 2D int array of shape (num_bboxes, 4), where each row represents a
bounding box: (y_top_left, x_top_left, box_height, box_width). Note that
containment is inclusive of the bounding box edges.
tap_distance_threshold: The threshold that determines if two taps result in
a matching screen state if they don't fall the same bounding boxes.
annotation_width_augment_fraction: The fraction to increase the width of the
bounding box by.
annotation_height_augment_fraction: The fraction to increase the height of
of the bounding box by.
Returns:
A boolean representing whether the two given actions are the same or not.
"""
action_1_touch_yx = jnp.asarray(action_1_touch_yx)
action_1_lift_yx = jnp.asarray(action_1_lift_yx)
action_2_touch_yx = jnp.asarray(action_2_touch_yx)
action_2_lift_yx = jnp.asarray(action_2_lift_yx)
# Checks if at least one of the actions is global (i.e. not DUAL_POINT),
# because if that is the case, only the actions' types need to be compared.
has_non_dual_point_action = jnp.logical_or(
_is_non_dual_point_action(action_1_action_type),
_is_non_dual_point_action(action_2_action_type),
)
#print("non dual point: "+str(has_non_dual_point_action))
different_dual_point_types = jnp.logical_xor(
is_tap_action(action_1_touch_yx, action_1_lift_yx),
is_tap_action(action_2_touch_yx, action_2_lift_yx),
)
#print("different dual type: "+str(different_dual_point_types))
is_tap = jnp.logical_and(
is_tap_action(action_1_touch_yx, action_1_lift_yx),
is_tap_action(action_2_touch_yx, action_2_lift_yx),
)
#print("is tap: "+str(is_tap))
taps_match = _check_tap_actions_match(
action_1_touch_yx,
action_2_touch_yx,
annotation_positions,
tap_distance_threshold,
annotation_width_augment_fraction,
annotation_height_augment_fraction,
)
#print("tap match: "+str(taps_match))
taps_match = jnp.logical_and(is_tap, taps_match)
#print("tap match: "+str(taps_match))
drags_match = _check_drag_actions_match(
action_1_touch_yx, action_1_lift_yx, action_2_touch_yx, action_2_lift_yx
)
drags_match = jnp.where(is_tap, False, drags_match)
#print("drag match: "+str(drags_match))
return jnp.where(
has_non_dual_point_action,
jnp.equal(action_1_action_type, action_2_action_type),
jnp.where(
different_dual_point_types,
False,
jnp.logical_or(taps_match, drags_match),
),
)
def action_2_format(step_data):
# 把test数据集中的动作格式转换为计算matching score的格式
action_type = step_data["action_type_id"]
if action_type == 4:
if step_data["action_type_text"] == 'click': # 点击
touch_point = step_data["touch"]
lift_point = step_data["lift"]
else: # 上下左右滑动
if step_data["action_type_text"] == 'scroll down':
touch_point = [0.5, 0.8]
lift_point = [0.5, 0.2]
elif step_data["action_type_text"] == 'scroll up':
touch_point = [0.5, 0.2]
lift_point = [0.5, 0.8]
elif step_data["action_type_text"] == 'scroll left':
touch_point = [0.2, 0.5]
lift_point = [0.8, 0.5]
elif step_data["action_type_text"] == 'scroll right':
touch_point = [0.8, 0.5]
lift_point = [0.2, 0.5]
else:
touch_point = [-1.0, -1.0]
lift_point = [-1.0, -1.0]
if action_type == 3:
typed_text = step_data["type_text"]
else:
typed_text = ""
action = {"action_type": action_type, "touch_point": touch_point, "lift_point": lift_point,
"typed_text": typed_text}
action["touch_point"] = [action["touch_point"][1], action["touch_point"][0]]
action["lift_point"] = [action["lift_point"][1], action["lift_point"][0]]
action["typed_text"] = action["typed_text"].lower()
return action
def pred_2_format(step_data):
# 把模型输出的内容转换为计算action_matching的格式
action_type = step_data["action_type"]
if action_type == 4: # 点击
action_type_new = 4
touch_point = step_data["click_point"]
lift_point = step_data["click_point"]
typed_text = ""
elif action_type == 0:
action_type_new = 4
touch_point = [0.5, 0.8]
lift_point = [0.5, 0.2]
typed_text = ""
elif action_type == 1:
action_type_new = 4
touch_point = [0.5, 0.2]
lift_point = [0.5, 0.8]
typed_text = ""
elif action_type == 8:
action_type_new = 4
touch_point = [0.2, 0.5]
lift_point = [0.8, 0.5]
typed_text = ""
elif action_type == 9:
action_type_new = 4
touch_point = [0.8, 0.5]
lift_point = [0.2, 0.5]
typed_text = ""
else:
action_type_new = action_type
touch_point = [-1.0, -1.0]
lift_point = [-1.0, -1.0]
typed_text = ""
if action_type_new == 3:
typed_text = step_data["typed_text"]
action = {"action_type": action_type_new, "touch_point": touch_point, "lift_point": lift_point,
"typed_text": typed_text}
action["touch_point"] = [action["touch_point"][1], action["touch_point"][0]]
action["lift_point"] = [action["lift_point"][1], action["lift_point"][0]]
action["typed_text"] = action["typed_text"].lower()
return action
def pred_2_format_simplified(step_data):
# 把模型输出的内容转换为计算action_matching的格式
action_type = step_data["action_type"]
if action_type == 'click' : # 点击
action_type_new = 4
touch_point = step_data["click_point"]
lift_point = step_data["click_point"]
typed_text = ""
elif action_type == 'scroll' and step_data["direction"] == 'down':
action_type_new = 4
touch_point = [0.5, 0.8]
lift_point = [0.5, 0.2]
typed_text = ""
elif action_type == 'scroll' and step_data["direction"] == 'up':
action_type_new = 4
touch_point = [0.5, 0.2]
lift_point = [0.5, 0.8]
typed_text = ""
elif action_type == 'scroll' and step_data["direction"] == 'left':
action_type_new = 4
touch_point = [0.2, 0.5]
lift_point = [0.8, 0.5]
typed_text = ""
elif action_type == 'scroll' and step_data["direction"] == 'right':
action_type_new = 4
touch_point = [0.8, 0.5]
lift_point = [0.2, 0.5]
typed_text = ""
elif action_type == 'type':
action_type_new = 3
touch_point = [-1.0, -1.0]
lift_point = [-1.0, -1.0]
typed_text = step_data["text"]
elif action_type == 'navigate_back':
action_type_new = 5
touch_point = [-1.0, -1.0]
lift_point = [-1.0, -1.0]
typed_text = ""
elif action_type == 'navigate_home':
action_type_new = 6
touch_point = [-1.0, -1.0]
lift_point = [-1.0, -1.0]
typed_text = ""
else:
action_type_new = action_type
touch_point = [-1.0, -1.0]
lift_point = [-1.0, -1.0]
typed_text = ""
# if action_type_new == 'type':
# typed_text = step_data["text"]
action = {"action_type": action_type_new, "touch_point": touch_point, "lift_point": lift_point,
"typed_text": typed_text}
action["touch_point"] = [action["touch_point"][1], action["touch_point"][0]]
action["lift_point"] = [action["lift_point"][1], action["lift_point"][0]]
action["typed_text"] = action["typed_text"].lower()
return action

45
util/action_type.py
View File

@@ -1,45 +0,0 @@
'''
Adapted from https://github.com/google-research/google-research/tree/master/android_in_the_wild
'''
import enum
class ActionType(enum.IntEnum):
# Placeholders for unused enum values
UNUSED_0 = 0
UNUSED_1 = 1
UNUSED_2 = 2
UNUSED_8 = 8
UNUSED_9 = 9
########### Agent actions ###########
# A type action that sends text to the emulator. Note that this simply sends
# text and does not perform any clicks for element focus or enter presses for
# submitting text.
TYPE = 3
# The dual point action used to represent all gestures.
DUAL_POINT = 4
# These actions differentiate pressing the home and back button from touches.
# They represent explicit presses of back and home performed using ADB.
PRESS_BACK = 5
PRESS_HOME = 6
# An action representing that ADB command for hitting enter was performed.
PRESS_ENTER = 7
########### Episode status actions ###########
# An action used to indicate the desired task has been completed and resets
# the environment. This action should also be used in the case that the task
# has already been completed and there is nothing to do.
# e.g. The task is to turn on the Wi-Fi when it is already on
STATUS_TASK_COMPLETE = 10
# An action used to indicate that desired task is impossible to complete and
# resets the environment. This can be a result of many different things
# including UI changes, Android version differences, etc.
STATUS_TASK_IMPOSSIBLE = 11

32
util/omniparser.py Normal file
View File

@@ -0,0 +1,32 @@
from util.utils import get_som_labeled_img, get_caption_model_processor, get_yolo_model, check_ocr_box
import torch
from PIL import Image
import io
import base64
from typing import Dict
class Omniparser(object):
def __init__(self, config: Dict):
self.config = config
device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.som_model = get_yolo_model(model_path=config['som_model_path'])
self.caption_model_processor = get_caption_model_processor(model_name=config['caption_model_name'], model_name_or_path=config['caption_model_path'], device=device)
print('Omniparser initialized!!!')
def parse(self, image_base64: str):
image_bytes = base64.b64decode(image_base64)
image = Image.open(io.BytesIO(image_bytes))
print('image size:', image.size)
box_overlay_ratio = max(image.size) / 3200
draw_bbox_config = {
'text_scale': 0.8 * box_overlay_ratio,
'text_thickness': max(int(2 * box_overlay_ratio), 1),
'text_padding': max(int(3 * box_overlay_ratio), 1),
'thickness': max(int(3 * box_overlay_ratio), 1),
}
(text, ocr_bbox), _ = check_ocr_box(image, display_img=False, output_bb_format='xyxy', easyocr_args={'text_threshold': 0.8}, use_paddleocr=False)
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(image, self.som_model, BOX_TRESHOLD = self.config['BOX_TRESHOLD'], output_coord_in_ratio=True, ocr_bbox=ocr_bbox,draw_bbox_config=draw_bbox_config, caption_model_processor=self.caption_model_processor, ocr_text=text,use_local_semantics=True, iou_threshold=0.7, scale_img=False, batch_size=128)
return dino_labled_img, parsed_content_list

108
utils.py → util/utils.py Executable file → Normal file
View File

@@ -35,12 +35,13 @@ import base64
import os
import ast
import torch
from typing import Tuple, List
from typing import Tuple, List, Union
from torchvision.ops import box_convert
import re
from torchvision.transforms import ToPILImage
import supervision as sv
import torchvision.transforms as T
from util.box_annotator import BoxAnnotator
def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
@@ -77,22 +78,21 @@ def get_yolo_model(model_path):
@torch.inference_mode()
def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=None):
# Number of samples per batch, --> 256 roughly takes 23 GB of GPU memory for florence model
to_pil = ToPILImage()
if starting_idx:
non_ocr_boxes = filtered_boxes[starting_idx:]
else:
non_ocr_boxes = filtered_boxes
croped_pil_image = []
t0 = time.time()
for i, coord in enumerate(non_ocr_boxes):
xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
cropped_image = image_source[ymin:ymax, xmin:xmax, :]
# resize the image to 224x224 to avoid long overhead in clipimageprocessor # TODO
cropped_image = cv2.resize(cropped_image, (224, 224))
croped_pil_image.append(to_pil(cropped_image))
print('time to prepare bbox:', time.time()-t0)
try:
xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
cropped_image = image_source[ymin:ymax, xmin:xmax, :]
cropped_image = cv2.resize(cropped_image, (64, 64))
croped_pil_image.append(to_pil(cropped_image))
except:
continue
model, processor = caption_model_processor['model'], caption_model_processor['processor']
if not prompt:
@@ -112,14 +112,10 @@ def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_
inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt", do_resize=False).to(device=device, dtype=torch.float16)
else:
inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device)
t2 = time.time()
print('time to process image + tokenize text inputs:', t2-t1)
if 'florence' in model.config.name_or_path:
generated_ids = model.generate(input_ids=inputs["input_ids"],pixel_values=inputs["pixel_values"],max_new_tokens=20,num_beams=1, do_sample=False)
else:
generated_ids = model.generate(**inputs, max_length=100, num_beams=5, no_repeat_ngram_size=2, early_stopping=True, num_return_sequences=1) # temperature=0.01, do_sample=True,
t3 = time.time()
print('time to generate:', t3-t2)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
generated_text = [gen.strip() for gen in generated_text]
generated_texts.extend(generated_text)
@@ -282,10 +278,10 @@ def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
is_valid_box = False
break
if is_valid_box:
# add the following 2 lines to include ocr bbox
if ocr_bbox:
# keep yolo boxes + prioritize ocr label
box_added = False
ocr_labels = ''
for box3_elem in ocr_bbox:
if not box_added:
box3 = box3_elem['bbox']
@@ -293,25 +289,22 @@ def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
# box_added = True
# delete the box3_elem from ocr_bbox
try:
filtered_boxes.append({'type': 'text', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': box3_elem['content']})
# gather all ocr labels
ocr_labels += box3_elem['content'] + ' '
filtered_boxes.remove(box3_elem)
# print('remove ocr bbox:', box3_elem)
except:
continue
# break
elif is_inside(box1, box3): # icon inside ocr
elif is_inside(box1, box3): # icon inside ocr, don't added this icon box, no need to check other ocr bbox bc no overlap between ocr bbox, icon can only be in one ocr box
box_added = True
# try:
# filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None})
# filtered_boxes.remove(box3_elem)
# except:
# continue
break
else:
continue
if not box_added:
filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None})
if ocr_labels:
filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': ocr_labels, 'source':'box_yolo_content_ocr'})
else:
filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None, 'source':'box_yolo_content_yolo'})
else:
filtered_boxes.append(box1)
return filtered_boxes # torch.tensor(filtered_boxes)
@@ -354,7 +347,6 @@ def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor
labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
from util.box_annotator import BoxAnnotator
box_annotator = BoxAnnotator(text_scale=text_scale, text_padding=text_padding,text_thickness=text_thickness,thickness=thickness) # 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
annotated_frame = image_source.copy()
annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w,h))
@@ -384,20 +376,20 @@ def predict(model, image, caption, box_threshold, text_threshold):
return boxes, logits, phrases
def predict_yolo(model, image_path, box_threshold, imgsz, scale_img, iou_threshold=0.7):
def predict_yolo(model, image, box_threshold, imgsz, scale_img, iou_threshold=0.7):
""" Use huggingface model to replace the original model
"""
# model = model['model']
if scale_img:
result = model.predict(
source=image_path,
source=image,
conf=box_threshold,
imgsz=imgsz,
iou=iou_threshold, # default 0.7
)
else:
result = model.predict(
source=image_path,
source=image,
conf=box_threshold,
iou=iou_threshold, # default 0.7
)
@@ -407,34 +399,41 @@ def predict_yolo(model, image_path, box_threshold, imgsz, scale_img, iou_thresho
return boxes, conf, phrases
def int_box_area(box, w, h):
x1, y1, x2, y2 = box
int_box = [int(x1*w), int(y1*h), int(x2*w), int(y2*h)]
area = (int_box[2] - int_box[0]) * (int_box[3] - int_box[1])
return area
def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD = 0.01, output_coord_in_ratio=False, ocr_bbox=None, text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None, ocr_text=[], use_local_semantics=True, iou_threshold=0.9,prompt=None, scale_img=False, imgsz=None, batch_size=64):
""" ocr_bbox: list of xyxy format bbox
def get_som_labeled_img(image_source: Union[str, Image.Image], model=None, BOX_TRESHOLD=0.01, output_coord_in_ratio=False, ocr_bbox=None, text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None, ocr_text=[], use_local_semantics=True, iou_threshold=0.9,prompt=None, scale_img=False, imgsz=None, batch_size=64):
"""Process either an image path or Image object
Args:
image_source: Either a file path (str) or PIL Image object
...
"""
image_source = Image.open(img_path).convert("RGB")
if isinstance(image_source, str):
image_source = Image.open(image_source).convert("RGB")
w, h = image_source.size
if not imgsz:
imgsz = (h, w)
# print('image size:', w, h)
xyxy, logits, phrases = predict_yolo(model=model, image_path=img_path, box_threshold=BOX_TRESHOLD, imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1)
xyxy, logits, phrases = predict_yolo(model=model, image=image_source, box_threshold=BOX_TRESHOLD, imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1)
xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
image_source = np.asarray(image_source)
phrases = [str(i) for i in range(len(phrases))]
# annotate the image with labels
h, w, _ = image_source.shape
if ocr_bbox:
ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
ocr_bbox=ocr_bbox.tolist()
else:
print('no ocr bbox!!!')
ocr_bbox = None
# filtered_boxes = remove_overlap(boxes=xyxy, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox)
# starting_idx = len(ocr_bbox)
# print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt} for box, txt in zip(ocr_bbox, ocr_text)]
xyxy_elem = [{'type': 'icon', 'bbox':box, 'interactivity':True, 'content':None} for box in xyxy.tolist()]
ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt, 'source': 'box_ocr_content_ocr'} for box, txt in zip(ocr_bbox, ocr_text) if int_box_area(box, w, h) > 0]
xyxy_elem = [{'type': 'icon', 'bbox':box, 'interactivity':True, 'content':None} for box in xyxy.tolist() if int_box_area(box, w, h) > 0]
filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
# sort the filtered_boxes so that the one with 'content': None is at the end, and get the index of the first 'content': None
@@ -444,7 +443,6 @@ def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD = 0.01, output_coord_
filtered_boxes = torch.tensor([box['bbox'] for box in filtered_boxes_elem])
print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
# get parsed icon local semantics
time1 = time.time()
if use_local_semantics:
@@ -483,7 +481,6 @@ def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD = 0.01, output_coord_
pil_img.save(buffered, format="PNG")
encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
if output_coord_in_ratio:
# h, w, _ = image_source.shape
label_coordinates = {k: [v[0]/w, v[1]/h, v[2]/w, v[3]/h] for k, v in label_coordinates.items()}
assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
@@ -504,46 +501,43 @@ def get_xywh_yolo(input):
x, y, w, h = input[0], input[1], input[2] - input[0], input[3] - input[1]
x, y, w, h = int(x), int(y), int(w), int(h)
return x, y, w, h
def check_ocr_box(image_path, display_img = True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
def check_ocr_box(image_source: Union[str, Image.Image], display_img = True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
if isinstance(image_source, str):
image_source = Image.open(image_source)
if image_source.mode == 'RGBA':
# Convert RGBA to RGB to avoid alpha channel issues
image_source = image_source.convert('RGB')
image_np = np.array(image_source)
w, h = image_source.size
if use_paddleocr:
if easyocr_args is None:
text_threshold = 0.5
else:
text_threshold = easyocr_args['text_threshold']
result = paddle_ocr.ocr(image_path, cls=False)[0]
# conf = [item[1] for item in result]
result = paddle_ocr.ocr(image_np, cls=False)[0]
coord = [item[0] for item in result if item[1][1] > text_threshold]
text = [item[1][0] for item in result if item[1][1] > text_threshold]
else: # EasyOCR
if easyocr_args is None:
easyocr_args = {}
result = reader.readtext(image_path, **easyocr_args)
# print('goal filtering pred:', result[-5:])
result = reader.readtext(image_np, **easyocr_args)
coord = [item[0] for item in result]
text = [item[1] for item in result]
# read the image using cv2
if display_img:
opencv_img = cv2.imread(image_path)
opencv_img = cv2.cvtColor(opencv_img, cv2.COLOR_RGB2BGR)
opencv_img = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
bb = []
for item in coord:
x, y, a, b = get_xywh(item)
# print(x, y, a, b)
bb.append((x, y, a, b))
cv2.rectangle(opencv_img, (x, y), (x+a, y+b), (0, 255, 0), 2)
# Display the image
plt.imshow(opencv_img)
# matplotlib expects RGB
plt.imshow(cv2.cvtColor(opencv_img, cv2.COLOR_BGR2RGB))
else:
if output_bb_format == 'xywh':
bb = [get_xywh(item) for item in coord]
elif output_bb_format == 'xyxy':
bb = [get_xyxy(item) for item in coord]
# print('bounding box!!!', bb)
return (text, bb), goal_filtering