feat(grounding): module smart_resize officiel Qwen3-VL

Module pur core/grounding/smart_resize.py implémentant la formule
smart_resize officielle (transformers.qwen2_vl.image_processing_qwen2_vl,
utilisée par Qwen3VLProcessor pour les images via wrap Qwen2VLImageProcessor).

Helpers exposés : _round_by_factor, _floor_by_factor, _ceil_by_factor.
Constantes : FACTOR_DEFAULT=28, MIN_PIXELS_DEFAULT=3136,
MAX_PIXELS_DEFAULT=1_003_520, MAX_RATIO_DEFAULT=200.

Tests : tests/unit/test_smart_resize.py — 32 cas, 100% coverage sur le
module (mesure via coverage API directe, pytest-cov bloqué par bug cv2
préexistant tracé dans DETTE-011).

refs DETTE-006 (étape 1/5 du fix smart_resize)
refs DETTE-007 (création de la 3ème implémentation, à unifier post-démo)
refs DETTE-010 (vérif preprocessor_config.json checkpoint Qwen3-VL-8B
                bloquante avant Étape 2)
refs DETTE-011 (bug cv2 contourné pour mesure coverage)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Dom
2026-05-09 12:42:47 +02:00
parent 4df1ba5779
commit 0d7bcd18ac
2 changed files with 311 additions and 0 deletions

View File

@@ -0,0 +1,77 @@
"""
Smart resize officiel Qwen3-VL (algorithme commun Qwen2-VL/Qwen3-VL pour images).
Source de référence : transformers.models.qwen2_vl.image_processing_qwen2_vl.smart_resize
(transformers 4.57.3). Qwen3-VL utilise Qwen2VLImageProcessor pour les images via
Qwen3VLProcessor.image_processor_class — la formule est donc commune Qwen2-VL/Qwen3-VL
sur le pipeline image.
Conditions garanties par smart_resize :
1. height et width retournés divisibles par `factor` (par défaut 28).
2. Total pixels dans l'intervalle [min_pixels, max_pixels].
3. Aspect ratio conservé au plus près.
Module image-only. Pour traitement vidéo Qwen3-VL (factor=32, autres bornes),
module dédié à créer si besoin futur.
"""
# DETTE-007 — Trois implémentations smart_resize coexistent dans le repo
# (core/grounding/server.py:15, core/grounding/infigui_worker.py:99, ce module).
# Unification post-démo Kerella.
import math
FACTOR_DEFAULT = 28
MIN_PIXELS_DEFAULT = 56 * 56 # 3136
MAX_PIXELS_DEFAULT = 14 * 14 * 4 * 1280 # 1_003_520
MAX_RATIO_DEFAULT = 200
def _round_by_factor(number: int, factor: int) -> int:
"""Closest integer to `number` divisible by `factor`."""
return round(number / factor) * factor
def _floor_by_factor(number: int, factor: int) -> int:
"""Largest integer ≤ `number` divisible by `factor`."""
return math.floor(number / factor) * factor
def _ceil_by_factor(number: int, factor: int) -> int:
"""Smallest integer ≥ `number` divisible by `factor`."""
return math.ceil(number / factor) * factor
def smart_resize(
height: int,
width: int,
factor: int = FACTOR_DEFAULT,
min_pixels: int = MIN_PIXELS_DEFAULT,
max_pixels: int = MAX_PIXELS_DEFAULT,
) -> tuple[int, int]:
"""Rescale (height, width) to satisfy the three conditions of the module docstring.
Raises:
ValueError: if max(height, width) / min(height, width) > MAX_RATIO_DEFAULT
(aspect ratio out of supported domain).
Returns:
(resized_height, resized_width).
"""
if max(height, width) / min(height, width) > MAX_RATIO_DEFAULT:
raise ValueError(
f"absolute aspect ratio must be smaller than {MAX_RATIO_DEFAULT}, "
f"got {max(height, width) / min(height, width)}"
)
h_bar = round(height / factor) * factor
w_bar = round(width / factor) * factor
if h_bar * w_bar > max_pixels:
beta = math.sqrt((height * width) / max_pixels)
h_bar = max(factor, math.floor(height / beta / factor) * factor)
w_bar = max(factor, math.floor(width / beta / factor) * factor)
elif h_bar * w_bar < min_pixels:
beta = math.sqrt(min_pixels / (height * width))
h_bar = math.ceil(height * beta / factor) * factor
w_bar = math.ceil(width * beta / factor) * factor
return h_bar, w_bar