EngineX-Cambricon/enginex-mlu370-any2any
9
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init

Browse Source
This commit is contained in:
luopingyi
2025-10-09 16:47:16 +08:00
parent c8feb4deb5
commit e27e3f16bb
5248 changed files with 1778505 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
transformers/tests/models/deit/__init__.py Normal file
View File

125
transformers/tests/models/deit/test_image_processing_deit.py Normal file
View File

@@ -0,0 +1,125 @@
# Copyright 2021 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 unittest
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torchvision_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
if is_vision_available():
from transformers import DeiTImageProcessor
if is_torchvision_available():
from transformers import DeiTImageProcessorFast
class DeiTImageProcessingTester:
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=400,
do_resize=True,
size=None,
do_center_crop=True,
crop_size=None,
do_normalize=True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
):
size = size if size is not None else {"height": 20, "width": 20}
crop_size = crop_size if crop_size is not None else {"height": 18, "width": 18}
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.min_resolution = min_resolution
self.max_resolution = max_resolution
self.do_resize = do_resize
self.size = size
self.do_center_crop = do_center_crop
self.crop_size = crop_size
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
def prepare_image_processor_dict(self):
return {
"do_resize": self.do_resize,
"size": self.size,
"do_center_crop": self.do_center_crop,
"crop_size": self.crop_size,
"do_normalize": self.do_normalize,
"image_mean": self.image_mean,
"image_std": self.image_std,
}
def expected_output_image_shape(self, images):
return self.num_channels, self.crop_size["height"], self.crop_size["width"]
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
return prepare_image_inputs(
batch_size=self.batch_size,
num_channels=self.num_channels,
min_resolution=self.min_resolution,
max_resolution=self.max_resolution,
equal_resolution=equal_resolution,
numpify=numpify,
torchify=torchify,
)
@require_torch
@require_vision
class DeiTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = DeiTImageProcessor if is_vision_available() else None
fast_image_processing_class = DeiTImageProcessorFast if is_torchvision_available() else None
test_cast_dtype = True
def setUp(self):
super().setUp()
self.image_processor_tester = DeiTImageProcessingTester(self)
@property
def image_processor_dict(self):
return self.image_processor_tester.prepare_image_processor_dict()
def test_image_processor_properties(self):
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
self.assertTrue(hasattr(image_processing, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "do_center_crop"))
self.assertTrue(hasattr(image_processing, "center_crop"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
def test_image_processor_from_dict_with_kwargs(self):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class.from_dict(self.image_processor_dict)
self.assertEqual(image_processor.size, {"height": 20, "width": 20})
self.assertEqual(image_processor.crop_size, {"height": 18, "width": 18})
image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42, crop_size=84)
self.assertEqual(image_processor.size, {"height": 42, "width": 42})
self.assertEqual(image_processor.crop_size, {"height": 84, "width": 84})

468
transformers/tests/models/deit/test_modeling_deit.py Normal file
View File

@@ -0,0 +1,468 @@
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# 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.
"""Testing suite for the PyTorch DeiT model."""
import unittest
import warnings
from functools import cached_property
from transformers import DeiTConfig
from transformers.testing_utils import (
require_accelerate,
require_torch,
require_torch_accelerator,
require_torch_fp16,
require_vision,
slow,
torch_device,
)
from transformers.utils import is_torch_available, is_vision_available
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import (
DeiTForImageClassification,
DeiTForImageClassificationWithTeacher,
DeiTForMaskedImageModeling,
DeiTModel,
)
from transformers.models.auto.modeling_auto import (
MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
MODEL_MAPPING_NAMES,
)
if is_vision_available():
from PIL import Image
from transformers import DeiTImageProcessor
class DeiTModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=30,
patch_size=2,
num_channels=3,
is_training=True,
use_labels=True,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
type_sequence_label_size=10,
initializer_range=0.02,
num_labels=3,
scope=None,
encoder_stride=2,
mask_ratio=0.5,
attn_implementation="eager",
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.is_training = is_training
self.use_labels = use_labels
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.scope = scope
self.encoder_stride = encoder_stride
self.attn_implementation = attn_implementation
# in DeiT, the seq length equals the number of patches + 2 (we add 2 for the [CLS] and distilation tokens)
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 2
self.mask_ratio = mask_ratio
self.num_masks = int(mask_ratio * self.seq_length)
self.mask_length = num_patches
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
labels = None
if self.use_labels:
labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return DeiTConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
is_decoder=False,
initializer_range=self.initializer_range,
encoder_stride=self.encoder_stride,
attn_implementation=self.attn_implementation,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DeiTModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_masked_image_modeling(self, config, pixel_values, labels):
model = DeiTForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.reconstruction.shape, (self.batch_size, self.num_channels, self.image_size, self.image_size)
)
# test greyscale images
config.num_channels = 1
model = DeiTForMaskedImageModeling(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.reconstruction.shape, (self.batch_size, 1, self.image_size, self.image_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.type_sequence_label_size
model = DeiTForImageClassification(config)
model.to(torch_device)
model.eval()
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = DeiTForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
pixel_values,
labels,
) = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_torch
class DeiTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as DeiT does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (
(
DeiTModel,
DeiTForImageClassification,
DeiTForImageClassificationWithTeacher,
DeiTForMaskedImageModeling,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = (
{
"image-feature-extraction": DeiTModel,
"image-classification": (DeiTForImageClassification, DeiTForImageClassificationWithTeacher),
}
if is_torch_available()
else {}
)
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
test_torch_exportable = True
def setUp(self):
self.model_tester = DeiTModelTester(self)
self.config_tester = ConfigTester(self, config_class=DeiTConfig, has_text_modality=False, hidden_size=37)
@unittest.skip(
"Since `torch==2.3+cu121`, although this test passes, many subsequent tests have `CUDA error: misaligned address`."
"If `nvidia-xxx-cu118` are also installed, no failure (even with `torch==2.3+cu121`)."
)
def test_multi_gpu_data_parallel_forward(self):
super().test_multi_gpu_data_parallel_forward()
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="DeiT does not use inputs_embeds")
def test_inputs_embeds(self):
pass
def test_model_get_set_embeddings(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_masked_image_modeling(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_masked_image_modeling(*config_and_inputs)
def test_for_image_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
# special case for DeiTForImageClassificationWithTeacher model
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
if return_labels:
if model_class.__name__ == "DeiTForImageClassificationWithTeacher":
del inputs_dict["labels"]
return inputs_dict
def test_training(self):
if not self.model_tester.is_training:
self.skipTest(reason="model_tester.is_training is set to False")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
for model_class in self.all_model_classes:
# DeiTForImageClassificationWithTeacher supports inference-only
if (
model_class.__name__ in MODEL_MAPPING_NAMES.values()
or model_class.__name__ == "DeiTForImageClassificationWithTeacher"
):
continue
model = model_class(config)
model.to(torch_device)
model.train()
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
loss = model(**inputs).loss
loss.backward()
def test_training_gradient_checkpointing(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if not self.model_tester.is_training:
self.skipTest(reason="model_tester.is_training is set to False")
config.use_cache = False
config.return_dict = True
for model_class in self.all_model_classes:
if model_class.__name__ in MODEL_MAPPING_NAMES.values() or not model_class.supports_gradient_checkpointing:
continue
# DeiTForImageClassificationWithTeacher supports inference-only
if model_class.__name__ == "DeiTForImageClassificationWithTeacher":
continue
model = model_class(config)
model.gradient_checkpointing_enable()
model.to(torch_device)
model.train()
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
loss = model(**inputs).loss
loss.backward()
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
def test_problem_types(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
problem_types = [
{"title": "multi_label_classification", "num_labels": 2, "dtype": torch.float},
{"title": "single_label_classification", "num_labels": 1, "dtype": torch.long},
{"title": "regression", "num_labels": 1, "dtype": torch.float},
]
for model_class in self.all_model_classes:
if (
model_class.__name__
not in [
*MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES.values(),
*MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES.values(),
]
or model_class.__name__ == "DeiTForImageClassificationWithTeacher"
):
continue
for problem_type in problem_types:
with self.subTest(msg=f"Testing {model_class} with {problem_type['title']}"):
config.problem_type = problem_type["title"]
config.num_labels = problem_type["num_labels"]
model = model_class(config)
model.to(torch_device)
model.train()
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
if problem_type["num_labels"] > 1:
inputs["labels"] = inputs["labels"].unsqueeze(1).repeat(1, problem_type["num_labels"])
inputs["labels"] = inputs["labels"].to(problem_type["dtype"])
# This tests that we do not trigger the warning form PyTorch "Using a target size that is different
# to the input size. This will likely lead to incorrect results due to broadcasting. Please ensure
# they have the same size." which is a symptom something in wrong for the regression problem.
# See https://github.com/huggingface/transformers/issues/11780
with warnings.catch_warnings(record=True) as warning_list:
loss = model(**inputs).loss
for w in warning_list:
if "Using a target size that is different to the input size" in str(w.message):
raise ValueError(
f"Something is going wrong in the regression problem: intercepted {w.message}"
)
loss.backward()
@slow
def test_model_from_pretrained(self):
model_name = "facebook/deit-base-distilled-patch16-224"
model = DeiTModel.from_pretrained(model_name)
self.assertIsNotNone(model)
# We will verify our results on an image of cute cats
def prepare_img():
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
return image
@require_torch
@require_vision
class DeiTModelIntegrationTest(unittest.TestCase):
@cached_property
def default_image_processor(self):
return (
DeiTImageProcessor.from_pretrained("facebook/deit-base-distilled-patch16-224")
if is_vision_available()
else None
)
@slow
def test_inference_image_classification_head(self):
model = DeiTForImageClassificationWithTeacher.from_pretrained("facebook/deit-base-distilled-patch16-224").to(
torch_device
)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the logits
expected_shape = torch.Size((1, 1000))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor([-1.0266, 0.1912, -1.2861]).to(torch_device)
torch.testing.assert_close(outputs.logits[0, :3], expected_slice, rtol=1e-4, atol=1e-4)
@slow
def test_inference_interpolate_pos_encoding(self):
model = DeiTForImageClassificationWithTeacher.from_pretrained("facebook/deit-base-distilled-patch16-224").to(
torch_device
)
image_processor = self.default_image_processor
# image size is {"height": 480, "width": 640}
image = prepare_img()
image_processor.size = {"height": 480, "width": 640}
# center crop set to False so image is not center cropped to 224x224
inputs = image_processor(images=image, return_tensors="pt", do_center_crop=False).to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs, interpolate_pos_encoding=True)
# verify the logits
expected_shape = torch.Size((1, 1000))
self.assertEqual(outputs.logits.shape, expected_shape)
@slow
@require_accelerate
@require_torch_accelerator
@require_torch_fp16
def test_inference_fp16(self):
r"""
A small test to make sure that inference work in half precision without any problem.
"""
model = DeiTModel.from_pretrained(
"facebook/deit-base-distilled-patch16-224", dtype=torch.float16, device_map="auto"
)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt")
pixel_values = inputs.pixel_values.to(torch_device)
# forward pass to make sure inference works in fp16
with torch.no_grad():
_ = model(pixel_values)