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_convnext/__init__.py Normal file
View File

243
transformers/tests/models/dinov3_convnext/test_modeling_dinov3_convnext.py Normal file
View File

@@ -0,0 +1,243 @@
# Copyright 2022 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 ConvNext model."""
import unittest
from functools import cached_property
from transformers import DINOv3ConvNextConfig
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, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import DINOv3ConvNextModel
if is_vision_available():
from PIL import Image
from transformers import AutoImageProcessor
class DINOv3ConvNextModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=32,
num_channels=3,
hidden_sizes=[10, 20, 30, 40],
depths=[2, 2, 3, 2],
is_training=False,
use_labels=True,
intermediate_size=37,
hidden_act="gelu",
num_labels=10,
initializer_range=0.02,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.num_channels = num_channels
self.hidden_sizes = hidden_sizes
self.depths = depths
self.is_training = is_training
self.use_labels = use_labels
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.num_labels = num_labels
self.initializer_range = initializer_range
self.scope = scope
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.num_labels)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return DINOv3ConvNextConfig(
num_channels=self.num_channels,
hidden_sizes=self.hidden_sizes,
depths=self.depths,
hidden_act=self.hidden_act,
is_decoder=False,
initializer_range=self.initializer_range,
num_labels=self.num_labels,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DINOv3ConvNextModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# expected last hidden states: B, C, H // 32, W // 32
self.parent.assertEqual(
result.last_hidden_state.shape,
(
self.batch_size,
1 + self.image_size // 32 * self.image_size // 32,
self.hidden_sizes[-1],
),
)
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 DINOv3ConvNextModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as ConvNext does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (DINOv3ConvNextModel,) if is_torch_available() else ()
pipeline_model_mapping = {"image-feature-extraction": DINOv3ConvNextModel} if is_torch_available() else {}
fx_compatible = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
has_attentions = False
test_torch_exportable = True
def setUp(self):
self.model_tester = DINOv3ConvNextModelTester(self)
self.config_tester = ConfigTester(
self,
config_class=DINOv3ConvNextConfig,
has_text_modality=False,
hidden_size=37,
common_properties=["num_channels", "hidden_sizes"],
)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="DINOv3ConvNext does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="DINOv3ConvNext does not support input and output embeddings")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="DINOv3ConvNext does not use feedforward chunking")
def test_feed_forward_chunking(self):
pass
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_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
self.assertEqual(len(hidden_states), 5)
# DINOv3ConvNext's feature maps are of shape (batch_size, num_channels, height, width)
self.assertListEqual(
list(hidden_states[1].shape[-2:]),
[self.model_tester.image_size // 4, self.model_tester.image_size // 4],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)
@slow
def test_model_from_pretrained(self):
model_name = "facebook/dinov3-convnext-tiny-pretrain-lvd1689m"
model = DINOv3ConvNextModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@unittest.skip(reason="DINOv3ConvNext does not retain grads for first hidden state (original pixel_values)")
def test_retain_grad_hidden_states_attentions(self):
pass
# 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 DINOv3ConvNextModelIntegrationTest(unittest.TestCase):
@cached_property
def default_image_processor(self):
return (
AutoImageProcessor.from_pretrained("facebook/dinov3-convnext-tiny-pretrain-lvd1689m")
if is_vision_available()
else None
)
@slow
def test_inference_no_head(self):
model = DINOv3ConvNextModel.from_pretrained("facebook/dinov3-convnext-tiny-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
_, _, height, width = inputs["pixel_values"].shape
expected_seq_length = (height * width) // 4 ** (model.config.num_stages + 1) + 1 # +1 for the "CLS" token
expected_shape = torch.Size((1, expected_seq_length, model.config.hidden_sizes[-1]))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
last_layer_cls_token = outputs.pooler_output
expected_slice = torch.tensor([-6.3721, 1.3008, 2.0743, -0.0800, 0.6072], 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[:, 1:]
expected_slice = torch.tensor([0.4905, -3.7135, 1.8485, -1.0403, -1.0908], device=torch_device)
torch.testing.assert_close(last_layer_patch_tokens[0, 0, :5], expected_slice, rtol=1e-4, atol=1e-4)