EngineX-Cambricon/enginex-mlu370-any2any
10
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/dinov3_vit/__init__.py Normal file
View File

127
transformers/tests/models/dinov3_vit/test_image_processing_dinov3_vit_fast.py Normal file
View File

@@ -0,0 +1,127 @@
# coding=utf-8
# Copyright 2024 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.
import unittest
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torchvision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
if is_torchvision_available():
from transformers import DINOv3ViTImageProcessorFast
class DINOv3ViTImageProcessingTester:
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=[0.48145466, 0.4578275, 0.40821073],
image_std=[0.26862954, 0.26130258, 0.27577711],
do_convert_rgb=True,
):
super().__init__()
size = size if size is not None else {"shortest_edge": 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
self.do_convert_rgb = do_convert_rgb
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,
"do_convert_rgb": self.do_convert_rgb,
}
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 DINOv3ViTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = None
fast_image_processing_class = DINOv3ViTImageProcessorFast if is_torchvision_available() else None
test_slow_image_processor = False
def setUp(self):
super().setUp()
self.image_processor_tester = DINOv3ViTImageProcessingTester(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"))
self.assertTrue(hasattr(image_processing, "do_convert_rgb"))
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, {"shortest_edge": 20})
self.assertEqual(image_processor.crop_size, {"height": 18, "width": 18})
image_processor = image_processing_class.from_dict(
self.image_processor_dict, size={"height": 42, "width": 42}
)
self.assertEqual(image_processor.size, {"height": 42, "width": 42})

279
transformers/tests/models/dinov3_vit/test_modeling_dinov3_vit.py Normal file
View File

@@ -0,0 +1,279 @@
# Copyright 2024 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 DINOv3 model."""
import unittest
from functools import cached_property
from transformers import DINOv3ViTConfig
from transformers.testing_utils import require_torch, 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, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import DINOv3ViTModel
if is_vision_available():
from PIL import Image
from transformers import AutoImageProcessor
class DINOv3ViTModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=30,
patch_size=2,
num_channels=3,
is_training=False,
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_register_tokens=2,
mask_ratio=0.5,
scope=None,
):
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.num_register_tokens = num_register_tokens
self.scope = scope
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 1 + self.num_register_tokens
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 DINOv3ViTConfig(
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,
num_register_tokens=self.num_register_tokens,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DINOv3ViTModel(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 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 Dinov3ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as Dinov3 does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (DINOv3ViTModel,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"image-feature-extraction": DINOv3ViTModel,
}
if is_torch_available()
else {}
)
fx_compatible = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
test_torch_exportable = True
def setUp(self):
self.model_tester = DINOv3ViTModelTester(self)
self.config_tester = ConfigTester(self, config_class=DINOv3ViTConfig, has_text_modality=False, hidden_size=37)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if param.requires_grad and "register_tokens" not in name:
# See PR #38607 (to avoid flakiness)
data = torch.flatten(param.data)
n_elements = torch.numel(data)
# skip 2.5% of elements on each side to avoid issues caused by `nn.init.trunc_normal_` described in
# https://github.com/huggingface/transformers/pull/27906#issuecomment-1846951332
n_elements_to_skip_on_each_side = int(n_elements * 0.025)
data_to_check = torch.sort(data).values
if n_elements_to_skip_on_each_side > 0:
data_to_check = data_to_check[n_elements_to_skip_on_each_side:-n_elements_to_skip_on_each_side]
self.assertIn(
((data_to_check.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="Dinov3 does not use inputs_embeds")
def test_inputs_embeds(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(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(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_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)
@unittest.skip(reason="Dinov3 does not support feedforward chunking yet")
def test_feed_forward_chunking(self):
pass
@slow
def test_model_from_pretrained(self):
model_name = "facebook/dinov3-vits16-pretrain-lvd1689m"
model = DINOv3ViTModel.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 DINOv3ViTModelIntegrationTest(unittest.TestCase):
@cached_property
def default_image_processor(self):
return (
AutoImageProcessor.from_pretrained("facebook/dinov3-vits16-pretrain-lvd1689m")
if is_vision_available()
else None
)
@slow
def test_inference_no_head(self):
model = DINOv3ViTModel.from_pretrained("facebook/dinov3-vits16-pretrain-lvd1689m").to(torch_device)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
# verify the last hidden states
# in DINOv3 with Registers, the seq length equals the number of patches + 1 + num_register_tokens (we add 1 for the [CLS] token)
_, _, height, width = inputs["pixel_values"].shape
num_patches = (height // model.config.patch_size) * (width // model.config.patch_size)
expected_seq_length = num_patches + 1 + model.config.num_register_tokens
expected_shape = torch.Size((1, expected_seq_length, model.config.hidden_size))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
last_layer_cls_token = outputs.pooler_output
expected_slice = torch.tensor([0.4637, -0.4160, 0.4086, -0.1265, -0.2865], device=torch_device)
torch.testing.assert_close(last_layer_cls_token[0, :5], expected_slice, rtol=1e-4, atol=1e-4)
last_layer_patch_tokens = outputs.last_hidden_state[:, model.config.num_register_tokens + 1 :]
expected_slice = torch.tensor([-0.0386, -0.2509, -0.0161, -0.4556, 0.5716], device=torch_device)
torch.testing.assert_close(last_layer_patch_tokens[0, 0, :5], expected_slice, rtol=1e-4, atol=1e-4)