SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

🚨[Fast Image Processor] Force Fast Image Processor for Qwen2_VL/2_5_VL + Refactor (#39591)

Browse Source 
* init

* Force qwen2VL image proc to fast

* refactor qwen2 vl fast

* fix copies

* Update after PR review and update tests to use return_tensors="pt"

* fix processor tests

* add BC for min pixels/max pixels
This commit is contained in:
Yoni Gozlan
2025-07-25 11:11:28 -04:00
committed by GitHub
parent f90de364c2
commit 17f02102c5
11 changed files with 222 additions and 283 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
tests/models/qwen2_5_omni/test_processor_qwen2_5_omni.py
View File

@@ -25,10 +25,17 @@ from huggingface_hub import hf_hub_download
from transformers import (
AutoProcessor,
Qwen2_5OmniProcessor,
Qwen2Tokenizer,
Qwen2TokenizerFast,
WhisperFeatureExtractor,
)
from transformers.testing_utils import require_av, require_librosa, require_torch, require_torchaudio, require_vision
from transformers.testing_utils import (
require_av,
require_librosa,
require_torch,
require_torchaudio,
require_torchvision,
require_vision,
)
from transformers.utils import is_torch_available, is_vision_available
from ...test_processing_common import ProcessorTesterMixin
@@ -38,12 +45,13 @@ if is_torch_available():
import torch
if is_vision_available():
from transformers import Qwen2VLImageProcessor
from transformers import Qwen2VLImageProcessorFast
@require_vision
@require_torch
@require_torchaudio
@require_torchvision
class Qwen2_5OmniProcessorTest(ProcessorTesterMixin, unittest.TestCase):
processor_class = Qwen2_5OmniProcessor
@@ -244,13 +252,13 @@ class Qwen2_5OmniProcessorTest(ProcessorTesterMixin, unittest.TestCase):
)
processor.save_pretrained(self.tmpdirname)
processor = Qwen2_5OmniProcessor.from_pretrained(self.tmpdirname, use_fast=False)
processor = Qwen2_5OmniProcessor.from_pretrained(self.tmpdirname, use_fast=True)
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
self.assertEqual(processor.image_processor.to_json_string(), image_processor.to_json_string())
self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
self.assertIsInstance(processor.tokenizer, Qwen2Tokenizer)
self.assertIsInstance(processor.image_processor, Qwen2VLImageProcessor)
self.assertIsInstance(processor.tokenizer, Qwen2TokenizerFast)
self.assertIsInstance(processor.image_processor, Qwen2VLImageProcessorFast)
self.assertIsInstance(processor.feature_extractor, WhisperFeatureExtractor)
def test_image_processor(self):
@@ -267,8 +275,8 @@ class Qwen2_5OmniProcessorTest(ProcessorTesterMixin, unittest.TestCase):
image_input = self.prepare_image_inputs()
input_image_proc = image_processor(image_input, return_tensors="np")
input_processor = processor(images=image_input, text="dummy", return_tensors="np")
input_image_proc = image_processor(image_input, return_tensors="pt")
input_processor = processor(images=image_input, text="dummy", return_tensors="pt")
for key in input_image_proc.keys():
self.assertAlmostEqual(input_image_proc[key].sum(), input_processor[key].sum(), delta=1e-2)

17
tests/models/qwen2_5_vl/test_processor_qwen2_5_vl.py
View File

@@ -20,15 +20,15 @@ import unittest
import numpy as np
import pytest
from transformers import AutoProcessor, Qwen2Tokenizer
from transformers.testing_utils import require_av, require_torch, require_vision
from transformers import AutoProcessor, Qwen2TokenizerFast
from transformers.testing_utils import require_av, require_torch, require_torchvision, require_vision
from transformers.utils import is_torch_available, is_vision_available
from ...test_processing_common import ProcessorTesterMixin
if is_vision_available():
from transformers import Qwen2_5_VLProcessor, Qwen2VLImageProcessor
from transformers import Qwen2_5_VLProcessor, Qwen2VLImageProcessorFast
if is_torch_available():
import torch
@@ -36,6 +36,7 @@ if is_torch_available():
@require_vision
@require_torch
@require_torchvision
class Qwen2_5_VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
processor_class = Qwen2_5_VLProcessor
@@ -73,12 +74,12 @@ class Qwen2_5_VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
tokenizer=tokenizer, image_processor=image_processor, video_processor=video_processor
)
processor.save_pretrained(self.tmpdirname)
processor = Qwen2_5_VLProcessor.from_pretrained(self.tmpdirname, use_fast=False)
processor = Qwen2_5_VLProcessor.from_pretrained(self.tmpdirname, use_fast=True)
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
self.assertEqual(processor.image_processor.to_json_string(), image_processor.to_json_string())
self.assertIsInstance(processor.tokenizer, Qwen2Tokenizer)
self.assertIsInstance(processor.image_processor, Qwen2VLImageProcessor)
self.assertIsInstance(processor.tokenizer, Qwen2TokenizerFast)
self.assertIsInstance(processor.image_processor, Qwen2VLImageProcessorFast)
def test_image_processor(self):
image_processor = self.get_image_processor()
@@ -91,8 +92,8 @@ class Qwen2_5_VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
image_input = self.prepare_image_inputs()
input_image_proc = image_processor(image_input, return_tensors="np")
input_processor = processor(images=image_input, text="dummy", return_tensors="np")
input_image_proc = image_processor(image_input, return_tensors="pt")
input_processor = processor(images=image_input, text="dummy", return_tensors="pt")
for key in input_image_proc.keys():
self.assertAlmostEqual(input_image_proc[key].sum(), input_processor[key].sum(), delta=1e-2)

39
tests/models/qwen2_vl/test_image_processing_qwen2_vl.py
View File

@@ -22,7 +22,7 @@ import requests
from transformers.image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_vision_available
from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs, prepare_video_inputs
@@ -35,8 +35,8 @@ if is_vision_available():
from transformers import Qwen2VLImageProcessor
# if is_torchvision_available():
# from transformers import Qwen2VLImageProcessorFast
if is_torchvision_available():
from transformers import Qwen2VLImageProcessorFast
class Qwen2VLImageProcessingTester:
@@ -119,7 +119,7 @@ class Qwen2VLImageProcessingTester:
@require_vision
class Qwen2VLImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = Qwen2VLImageProcessor if is_vision_available() else None
# fast_image_processing_class = Qwen2VLImageProcessorFast if is_torchvision_available() else None
fast_image_processing_class = Qwen2VLImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
@@ -363,3 +363,34 @@ class Qwen2VLImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
encoding_fast = image_processor_fast(dummy_image, return_tensors="pt")
self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values, encoding_fast.pixel_values)
self.assertEqual(encoding_slow.image_grid_thw.dtype, encoding_fast.image_grid_thw.dtype)
self._assert_slow_fast_tensors_equivalence(
encoding_slow.image_grid_thw.float(), encoding_fast.image_grid_thw.float()
)
@require_vision
@require_torch
def test_slow_fast_equivalence_batched(self):
if not self.test_slow_image_processor or not self.test_fast_image_processor:
self.skipTest(reason="Skipping slow/fast equivalence test")
if self.image_processing_class is None or self.fast_image_processing_class is None:
self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
self.skipTest(
reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
)
dummy_images = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
image_processor_slow = self.image_processing_class(**self.image_processor_dict)
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
encoding_slow = image_processor_slow(dummy_images, return_tensors="pt")
encoding_fast = image_processor_fast(dummy_images, return_tensors="pt")
self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values, encoding_fast.pixel_values)
self.assertEqual(encoding_slow.image_grid_thw.dtype, encoding_fast.image_grid_thw.dtype)
self._assert_slow_fast_tensors_equivalence(
encoding_slow.image_grid_thw.float(), encoding_fast.image_grid_thw.float()
)

19
tests/models/qwen2_vl/test_processor_qwen2_vl.py
View File

@@ -20,18 +20,18 @@ import unittest
import numpy as np
import pytest
from transformers import AutoProcessor, Qwen2Tokenizer
from transformers.testing_utils import require_av, require_torch, require_vision
from transformers import AutoProcessor, Qwen2TokenizerFast
from transformers.testing_utils import require_av, require_torch, require_torchvision, require_vision
from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
from ...test_processing_common import ProcessorTesterMixin
if is_vision_available():
from transformers import Qwen2VLImageProcessor, Qwen2VLProcessor
from transformers import Qwen2VLProcessor
if is_torchvision_available():
from transformers import Qwen2VLVideoProcessor
from transformers import Qwen2VLImageProcessorFast, Qwen2VLVideoProcessor
if is_torch_available():
import torch
@@ -39,6 +39,7 @@ if is_torch_available():
@require_vision
@require_torch
@require_torchvision
class Qwen2VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
processor_class = Qwen2VLProcessor
@@ -76,12 +77,12 @@ class Qwen2VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
tokenizer=tokenizer, image_processor=image_processor, video_processor=video_processor
)
processor.save_pretrained(self.tmpdirname)
processor = Qwen2VLProcessor.from_pretrained(self.tmpdirname, use_fast=False)
processor = Qwen2VLProcessor.from_pretrained(self.tmpdirname, use_fast=True)
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
self.assertEqual(processor.image_processor.to_json_string(), image_processor.to_json_string())
self.assertIsInstance(processor.tokenizer, Qwen2Tokenizer)
self.assertIsInstance(processor.image_processor, Qwen2VLImageProcessor)
self.assertIsInstance(processor.tokenizer, Qwen2TokenizerFast)
self.assertIsInstance(processor.image_processor, Qwen2VLImageProcessorFast)
self.assertIsInstance(processor.video_processor, Qwen2VLVideoProcessor)
def test_image_processor(self):
@@ -95,8 +96,8 @@ class Qwen2VLProcessorTest(ProcessorTesterMixin, unittest.TestCase):
image_input = self.prepare_image_inputs()
input_image_proc = image_processor(image_input, return_tensors="np")
input_processor = processor(images=image_input, text="dummy", return_tensors="np")
input_image_proc = image_processor(image_input, return_tensors="pt")
input_processor = processor(images=image_input, text="dummy", return_tensors="pt")
for key in input_image_proc.keys():
self.assertAlmostEqual(input_image_proc[key].sum(), input_processor[key].sum(), delta=1e-2)

2
tests/test_processing_common.py
View File

@@ -937,7 +937,7 @@ class ProcessorTesterMixin:
"video", batch_size, return_tensors, "videos_input_name", "video_processor", MODALITY_INPUT_DATA["videos"]
)
@parameterized.expand([(1, "np"), (1, "pt"), (2, "np"), (2, "pt")])
@parameterized.expand([(1, "pt"), (2, "pt")]) # fast image processors supports only torchvision
def test_apply_chat_template_image(self, batch_size: int, return_tensors: str):
self._test_apply_chat_template(
"image", batch_size, return_tensors, "images_input_name", "image_processor", MODALITY_INPUT_DATA["images"]