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

Uniformize LlavaNextVideoProcessor kwargs (#35613)

Browse Source 
* Uniformize processor kwargs and add tests

* add videos_kwargs tests

* fix copies

* fix llava_next_video chat template tests

* remove unnecessary default kwargs
This commit is contained in:
Yoni Gozlan
2025-02-18 14:13:51 -05:00
committed by GitHub
parent 8ee50537fe
commit 9b479a245b
5 changed files with 402 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

166
tests/models/llava_next_video/test_processor_llava_next_video.py Normal file
View File

@@ -0,0 +1,166 @@
# coding=utf-8
# Copyright 2025 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import json
import shutil
import tempfile
import unittest
from transformers import AutoProcessor, LlamaTokenizerFast, LlavaNextVideoProcessor
from transformers.testing_utils import require_av, require_torch, 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 LlavaNextImageProcessor, LlavaNextVideoImageProcessor
if is_torch_available:
import torch
@require_vision
class LlavaNextVideoProcessorTest(ProcessorTesterMixin, unittest.TestCase):
processor_class = LlavaNextVideoProcessor
def setUp(self):
self.tmpdirname = tempfile.mkdtemp()
image_processor = LlavaNextImageProcessor()
video_processor = LlavaNextVideoImageProcessor()
tokenizer = LlamaTokenizerFast.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
processor_kwargs = self.prepare_processor_dict()
processor = LlavaNextVideoProcessor(
video_processor=video_processor, image_processor=image_processor, tokenizer=tokenizer, **processor_kwargs
)
processor.save_pretrained(self.tmpdirname)
def get_tokenizer(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
def get_image_processor(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
def get_video_processor(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).video_processor
def prepare_processor_dict(self):
return {
"chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + ' '}}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>' }}{% endfor %}{# Render all video then #}{% for content in message['content'] | selectattr('type', 'equalto', 'video') %}{{ '<video>' }}{% endfor %}{# Render all text next #}{% if message['role'] != 'assistant' %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ '\n' + content['text'] }}{% endfor %}{% else %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{% generation %}{{ '\n' + content['text'] }}{% endgeneration %}{% endfor %}{% endif %}{{'<|im_end|>'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
"num_additional_image_tokens": 6,
"patch_size": 4,
"vision_feature_select_strategy": "default",
}
def test_processor_to_json_string(self):
processor = self.get_processor()
obj = json.loads(processor.to_json_string())
for key, value in self.prepare_processor_dict().items():
# chat_tempalate are tested as a separate test because they are saved in separate files
if key != "chat_template":
self.assertEqual(obj[key], value)
self.assertEqual(getattr(processor, key, None), value)
# Copied from tests.models.llava.test_processor_llava.LlavaProcessorTest.test_chat_template_is_saved
def test_chat_template_is_saved(self):
processor_loaded = self.processor_class.from_pretrained(self.tmpdirname)
processor_dict_loaded = json.loads(processor_loaded.to_json_string())
# chat templates aren't serialized to json in processors
self.assertFalse("chat_template" in processor_dict_loaded.keys())
# they have to be saved as separate file and loaded back from that file
# so we check if the same template is loaded
processor_dict = self.prepare_processor_dict()
self.assertTrue(processor_loaded.chat_template == processor_dict.get("chat_template", None))
def tearDown(self):
shutil.rmtree(self.tmpdirname)
def test_chat_template(self):
processor = AutoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
expected_prompt = "USER: <image>\nWhat is shown in this image? ASSISTANT:"
messages = [
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": "What is shown in this image?"},
],
},
]
formatted_prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
self.assertEqual(expected_prompt, formatted_prompt)
@require_av
def test_chat_template_dict(self):
processor = AutoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
messages = [
{
"role": "user",
"content": [
{"type": "video"},
{"type": "text", "text": "What is shown in this video?"},
],
},
]
formatted_prompt_tokenized = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_tensors=None
)
expected_output = [[1, 3148, 1001, 29901, 29871, 32000, 13, 5618, 338, 4318, 297, 445, 4863, 29973, 319, 1799, 9047, 13566, 29901]] # fmt: skip
self.assertListEqual(expected_output, formatted_prompt_tokenized)
out_dict = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True)
self.assertListEqual(list(out_dict.keys()), ["input_ids", "attention_mask"])
# add image URL for return dict
messages[0]["content"][0] = {
"type": "video",
"url": "https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/720/Big_Buck_Bunny_720_10s_10MB.mp4",
}
out_dict_with_video = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_dict=True
)
self.assertListEqual(list(out_dict_with_video.keys()), ["input_ids", "attention_mask", "pixel_values_videos"])
@require_torch
@require_av
def test_chat_template_dict_torch(self):
processor = AutoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
messages = [
{
"role": "user",
"content": [
{
"type": "video",
"url": "https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/720/Big_Buck_Bunny_720_10s_10MB.mp4",
},
{"type": "text", "text": "What is shown in this video?"},
],
},
]
out_dict_tensors = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
)
self.assertListEqual(list(out_dict_tensors.keys()), ["input_ids", "attention_mask", "pixel_values_videos"])
self.assertTrue(isinstance(out_dict_tensors["input_ids"], torch.Tensor))

209
tests/test_processing_common.py
View File

@@ -126,11 +126,12 @@ class ProcessorTesterMixin:
return prepare_image_inputs() * batch_size
@require_vision
def prepare_video_inputs(self):
def prepare_video_inputs(self, batch_size: Optional[int] = None):
"""This function prepares a list of numpy videos."""
video_input = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)] * 8
image_inputs = [video_input] * 3 # batch-size=3
return image_inputs
if batch_size is None:
return video_input
return [video_input] * batch_size
def test_processor_to_json_string(self):
processor = self.get_processor()
@@ -491,6 +492,192 @@ class ProcessorTesterMixin:
elif "labels" in inputs:
self.assertEqual(len(inputs["labels"][0]), 76)
def test_tokenizer_defaults_preserved_by_kwargs_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
inputs = processor(text=input_str, videos=video_input, return_tensors="pt")
self.assertEqual(inputs[self.text_input_name].shape[-1], 117)
def test_video_processor_defaults_preserved_by_video_kwargs(self):
"""
We use do_rescale=True, rescale_factor=-1 to ensure that image_processor kwargs are preserved in the processor.
We then check that the mean of the pixel_values is less than or equal to 0 after processing.
Since the original pixel_values are in [0, 255], this is a good indicator that the rescale_factor is indeed applied.
"""
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_components["video_processor"] = self.get_component(
"video_processor", do_rescale=True, rescale_factor=-1
)
processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
inputs = processor(text=input_str, videos=video_input, return_tensors="pt")
self.assertLessEqual(inputs[self.videos_input_name][0][0][0].mean(), 0)
def test_kwargs_overrides_default_tokenizer_kwargs_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_components["tokenizer"] = self.get_component("tokenizer", padding="longest")
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
inputs = processor(
text=input_str, videos=video_input, return_tensors="pt", max_length=112, padding="max_length"
)
self.assertEqual(inputs[self.text_input_name].shape[-1], 112)
def test_kwargs_overrides_default_video_processor_kwargs(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_components["video_processor"] = self.get_component(
"video_processor", do_rescale=True, rescale_factor=1
)
processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
inputs = processor(text=input_str, videos=video_input, do_rescale=True, rescale_factor=-1, return_tensors="pt")
self.assertLessEqual(inputs[self.videos_input_name][0][0][0].mean(), 0)
def test_unstructured_kwargs_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
inputs = processor(
text=input_str,
videos=video_input,
return_tensors="pt",
do_rescale=True,
rescale_factor=-1,
padding="max_length",
max_length=76,
)
self.assertLessEqual(inputs[self.videos_input_name][0][0][0].mean(), 0)
self.assertEqual(inputs[self.text_input_name].shape[-1], 76)
def test_unstructured_kwargs_batched_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs(batch_size=2)
video_input = self.prepare_video_inputs(batch_size=2)
inputs = processor(
text=input_str,
videos=video_input,
return_tensors="pt",
do_rescale=True,
rescale_factor=-1,
padding="longest",
max_length=76,
)
self.assertLessEqual(inputs[self.videos_input_name][0][0][0].mean(), 0)
self.assertTrue(
len(inputs[self.text_input_name][0]) == len(inputs[self.text_input_name][1])
and len(inputs[self.text_input_name][1]) < 76
)
def test_doubly_passed_kwargs_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = [self.prepare_text_inputs()]
video_input = self.prepare_video_inputs()
with self.assertRaises(ValueError):
_ = processor(
text=input_str,
videos=video_input,
videos_kwargs={"do_rescale": True, "rescale_factor": -1},
do_rescale=True,
return_tensors="pt",
)
def test_structured_kwargs_nested_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
# Define the kwargs for each modality
all_kwargs = {
"common_kwargs": {"return_tensors": "pt"},
"videos_kwargs": {"do_rescale": True, "rescale_factor": -1},
"text_kwargs": {"padding": "max_length", "max_length": 76},
}
inputs = processor(text=input_str, videos=video_input, **all_kwargs)
self.skip_processor_without_typed_kwargs(processor)
self.assertLessEqual(inputs[self.videos_input_name][0][0][0].mean(), 0)
self.assertEqual(inputs[self.text_input_name].shape[-1], 76)
def test_structured_kwargs_nested_from_dict_video(self):
if "video_processor" not in self.processor_class.attributes:
self.skipTest(f"video_processor attribute not present in {self.processor_class}")
processor_components = self.prepare_components()
processor_kwargs = self.prepare_processor_dict()
processor = self.processor_class(**processor_components, **processor_kwargs)
self.skip_processor_without_typed_kwargs(processor)
input_str = self.prepare_text_inputs()
video_input = self.prepare_video_inputs()
# Define the kwargs for each modality
all_kwargs = {
"common_kwargs": {"return_tensors": "pt"},
"videos_kwargs": {"do_rescale": True, "rescale_factor": -1},
"text_kwargs": {"padding": "max_length", "max_length": 76},
}
inputs = processor(text=input_str, videos=video_input, **all_kwargs)
self.assertLessEqual(inputs[self.videos_input_name][0][0][0].mean(), 0)
self.assertEqual(inputs[self.text_input_name].shape[-1], 76)
# TODO: the same test, but for audio + text processors that have strong overlap in kwargs
# TODO (molbap) use the same structure of attribute kwargs for other tests to avoid duplication
def test_overlapping_text_kwargs_handling(self):
@@ -584,7 +771,9 @@ class ProcessorTesterMixin:
formatted_prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
self.assertEqual(len(formatted_prompt), 1)
formatted_prompt_tokenized = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True)
formatted_prompt_tokenized = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_tensors=None
)
add_special_tokens = True
if processor.tokenizer.bos_token is not None and formatted_prompt[0].startswith(processor.tokenizer.bos_token):
add_special_tokens = False
@@ -636,7 +825,7 @@ class ProcessorTesterMixin:
self.assertEqual(len(formatted_prompt), 2)
formatted_prompt_tokenized = processor.apply_chat_template(
batched_messages, add_generation_prompt=True, tokenize=True, padding=True
batched_messages, add_generation_prompt=True, tokenize=True, padding=True, return_tensors=None
)
add_special_tokens = True
if processor.tokenizer.bos_token is not None and formatted_prompt[0].startswith(processor.tokenizer.bos_token):
@@ -650,7 +839,11 @@ class ProcessorTesterMixin:
self.assertListEqual(expected_output, formatted_prompt_tokenized)
out_dict = processor.apply_chat_template(
batched_messages, add_generation_prompt=True, tokenize=True, return_dict=True, padding=True
batched_messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
padding=True,
)
self.assertListEqual(list(out_dict.keys()), ["input_ids", "attention_mask"])
@@ -775,7 +968,9 @@ class ProcessorTesterMixin:
formatted_prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
self.assertEqual(len(formatted_prompt), 1)
formatted_prompt_tokenized = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True)
formatted_prompt_tokenized = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_tensors=None
)
add_special_tokens = True
if processor.tokenizer.bos_token is not None and formatted_prompt[0].startswith(processor.tokenizer.bos_token):
add_special_tokens = False