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/dpt/__init__.py Normal file
View File

353
transformers/tests/models/dpt/test_image_processing_dpt.py Normal file
View File

@@ -0,0 +1,353 @@
# Copyright 2022 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
import numpy as np
from datasets import load_dataset
from transformers.file_utils import is_torch_available, is_vision_available
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_torch_available():
import torch
if is_vision_available():
from transformers import DPTImageProcessor
if is_torchvision_available():
from transformers import DPTImageProcessorFast
class DPTImageProcessingTester:
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_normalize=True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
do_reduce_labels=False,
):
size = size if 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_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
self.do_reduce_labels = do_reduce_labels
def prepare_image_processor_dict(self):
return {
"image_mean": self.image_mean,
"image_std": self.image_std,
"do_normalize": self.do_normalize,
"do_resize": self.do_resize,
"size": self.size,
"do_reduce_labels": self.do_reduce_labels,
}
def expected_output_image_shape(self, images):
return self.num_channels, self.size["height"], self.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,
)
# Copied from transformers.tests.models.beit.test_image_processing_beit.prepare_semantic_single_inputs
def prepare_semantic_single_inputs():
ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
example = ds[0]
return example["image"], example["map"]
# Copied from transformers.tests.models.beit.test_image_processing_beit.prepare_semantic_batch_inputs
def prepare_semantic_batch_inputs():
ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
return list(ds["image"][:2]), list(ds["map"][:2])
@require_torch
@require_vision
class DPTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = DPTImageProcessor if is_vision_available() else None
fast_image_processing_class = DPTImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = DPTImageProcessingTester(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, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "do_rescale"))
self.assertTrue(hasattr(image_processing, "rescale_factor"))
self.assertTrue(hasattr(image_processing, "do_pad"))
self.assertTrue(hasattr(image_processing, "size_divisor"))
self.assertTrue(hasattr(image_processing, "do_reduce_labels"))
def test_image_processor_from_dict_with_kwargs(self):
for image_processing_class in self.image_processor_list:
image_processing_class = image_processing_class(**self.image_processor_dict)
image_processor = image_processing_class.from_dict(self.image_processor_dict)
self.assertEqual(image_processor.size, {"height": 18, "width": 18})
image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42)
self.assertEqual(image_processor.size, {"height": 42, "width": 42})
def test_padding(self):
for image_processing_class in self.image_processor_list:
if image_processing_class == DPTImageProcessorFast:
image = torch.arange(0, 366777, 1, dtype=torch.uint8).reshape(3, 249, 491)
image_processor = image_processing_class(**self.image_processor_dict)
padded_image = image_processor.pad_image(image, size_divisor=4)
self.assertTrue(padded_image.shape[1] % 4 == 0)
self.assertTrue(padded_image.shape[2] % 4 == 0)
pixel_values = image_processor.preprocess(
image, do_rescale=False, do_resize=False, do_pad=True, size_divisor=4, return_tensors="pt"
).pixel_values
self.assertTrue(pixel_values.shape[2] % 4 == 0)
self.assertTrue(pixel_values.shape[3] % 4 == 0)
else:
image_processor = image_processing_class(**self.image_processor_dict)
image = np.random.randn(3, 249, 491)
image = image_processor.pad_image(image, size_divisor=4)
self.assertTrue(image.shape[1] % 4 == 0)
self.assertTrue(image.shape[2] % 4 == 0)
pixel_values = image_processor.preprocess(
image, do_rescale=False, do_resize=False, do_pad=True, size_divisor=4, return_tensors="pt"
).pixel_values
self.assertTrue(pixel_values.shape[2] % 4 == 0)
self.assertTrue(pixel_values.shape[3] % 4 == 0)
def test_keep_aspect_ratio(self):
size = {"height": 512, "width": 512}
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(size=size, keep_aspect_ratio=True, ensure_multiple_of=32)
image = np.zeros((489, 640, 3))
pixel_values = image_processor(image, return_tensors="pt").pixel_values
self.assertEqual(list(pixel_values.shape), [1, 3, 512, 672])
# Copied from transformers.tests.models.beit.test_image_processing_beit.BeitImageProcessingTest.test_call_segmentation_maps
def test_call_segmentation_maps(self):
for image_processing_class in self.image_processor_list:
# Initialize image_processor
image_processor = image_processing_class(**self.image_processor_dict)
# create random PyTorch tensors
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
maps = []
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
maps.append(torch.zeros(image.shape[-2:]).long())
# Test not batched input
encoding = image_processor(image_inputs[0], maps[0], return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
1,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
1,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Test batched
encoding = image_processor(image_inputs, maps, return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Test not batched input (PIL images)
image, segmentation_map = prepare_semantic_single_inputs()
encoding = image_processor(image, segmentation_map, return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
1,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
1,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Test batched input (PIL images)
images, segmentation_maps = prepare_semantic_batch_inputs()
encoding = image_processor(images, segmentation_maps, return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
2,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
2,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
def test_reduce_labels(self):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
# ADE20k has 150 classes, and the background is included, so labels should be between 0 and 150
image, map = prepare_semantic_single_inputs()
encoding = image_processor(image, map, return_tensors="pt")
labels_no_reduce = encoding["labels"].clone()
self.assertTrue(labels_no_reduce.min().item() >= 0)
self.assertTrue(labels_no_reduce.max().item() <= 150)
# Get the first non-zero label coords and value, for comparison when do_reduce_labels is True
non_zero_positions = (labels_no_reduce > 0).nonzero()
first_non_zero_coords = tuple(non_zero_positions[0].tolist())
first_non_zero_value = labels_no_reduce[first_non_zero_coords].item()
image_processor.do_reduce_labels = True
encoding = image_processor(image, map, return_tensors="pt")
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Compare with non-reduced label to see if it's reduced by 1
self.assertEqual(encoding["labels"][first_non_zero_coords].item(), first_non_zero_value - 1)
def test_slow_fast_equivalence(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")
dummy_image, dummy_map = prepare_semantic_single_inputs()
image_processor_slow = self.image_processing_class(**self.image_processor_dict)
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
image_encoding_slow = image_processor_slow(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
image_encoding_fast = image_processor_fast(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
self.assertTrue(torch.allclose(image_encoding_slow.pixel_values, image_encoding_fast.pixel_values, atol=1e-1))
self.assertLessEqual(
torch.mean(torch.abs(image_encoding_slow.pixel_values - image_encoding_fast.pixel_values)).item(), 1e-3
)
self.assertTrue(torch.allclose(image_encoding_slow.labels, image_encoding_fast.labels, atol=1e-1))
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, dummy_maps = prepare_semantic_batch_inputs()
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, segmentation_maps=dummy_maps, return_tensors="pt")
encoding_fast = image_processor_fast(dummy_images, segmentation_maps=dummy_maps, return_tensors="pt")
self.assertTrue(torch.allclose(encoding_slow.pixel_values, encoding_fast.pixel_values, atol=1e-1))
self.assertLessEqual(
torch.mean(torch.abs(encoding_slow.pixel_values - encoding_fast.pixel_values)).item(), 1e-3
)

446
transformers/tests/models/dpt/test_modeling_dpt.py Normal file
View File

@@ -0,0 +1,446 @@
# 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 DPT model."""
import unittest
import pytest
from transformers import DPTConfig
from transformers.file_utils import is_torch_available, is_vision_available
from transformers.pytorch_utils import is_torch_greater_or_equal_than_2_4
from transformers.testing_utils import Expectations, require_torch, require_vision, slow, torch_device
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 DPTForDepthEstimation, DPTForSemanticSegmentation, DPTModel
from transformers.models.auto.modeling_auto import MODEL_MAPPING_NAMES
if is_vision_available():
from PIL import Image
from transformers import DPTImageProcessor
class DPTModelTester:
def __init__(
self,
parent,
batch_size=2,
image_size=32,
patch_size=16,
num_channels=3,
is_training=True,
use_labels=True,
hidden_size=32,
num_hidden_layers=2,
backbone_out_indices=[0, 1, 2, 3],
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
initializer_range=0.02,
num_labels=3,
neck_hidden_sizes=[16, 32],
is_hybrid=False,
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.backbone_out_indices = backbone_out_indices
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.initializer_range = initializer_range
self.num_labels = num_labels
self.scope = scope
self.is_hybrid = is_hybrid
self.neck_hidden_sizes = neck_hidden_sizes
# sequence length of DPT = num_patches + 1 (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 1
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.image_size, self.image_size], self.num_labels)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return DPTConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
fusion_hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
backbone_out_indices=self.backbone_out_indices,
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,
is_hybrid=self.is_hybrid,
neck_hidden_sizes=self.neck_hidden_sizes,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DPTModel(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_depth_estimation(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = DPTForDepthEstimation(config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(result.predicted_depth.shape, (self.batch_size, self.image_size, self.image_size))
def create_and_check_for_semantic_segmentation(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = DPTForSemanticSegmentation(config)
model.to(torch_device)
model.eval()
result = model(pixel_values, labels=labels)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_labels, self.image_size, self.image_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 DPTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as DPT does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (DPTModel, DPTForDepthEstimation, DPTForSemanticSegmentation) if is_torch_available() else ()
pipeline_model_mapping = (
{
"depth-estimation": DPTForDepthEstimation,
"image-feature-extraction": DPTModel,
"image-segmentation": DPTForSemanticSegmentation,
}
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 = DPTModelTester(self)
self.config_tester = ConfigTester(self, config_class=DPTConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="DPT 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_depth_estimation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_depth_estimation(*config_and_inputs)
def test_for_semantic_segmentation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_semantic_segmentation(*config_and_inputs)
def test_training(self):
for model_class in self.all_model_classes:
if model_class.__name__ == "DPTForDepthEstimation":
continue
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
if model_class.__name__ in MODEL_MAPPING_NAMES.values():
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):
for model_class in self.all_model_classes:
if model_class.__name__ == "DPTForDepthEstimation":
continue
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.use_cache = False
config.return_dict = True
if model_class.__name__ in MODEL_MAPPING_NAMES.values() or not model_class.supports_gradient_checkpointing:
continue
model = model_class(config)
model.to(torch_device)
model.gradient_checkpointing_enable()
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
@unittest.skip(reason="Inductor error for dynamic shape")
@pytest.mark.torch_compile_test
def test_sdpa_can_compile_dynamic(self):
pass
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)
# Skip the check for the backbone
backbone_params = []
for name, module in model.named_modules():
if module.__class__.__name__ == "DPTViTHybridEmbeddings":
backbone_params = [f"{name}.{key}" for key in module.state_dict()]
break
for name, param in model.named_parameters():
if param.requires_grad:
if name in backbone_params:
continue
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
def test_backbone_selection(self):
def _validate_backbone_init():
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
if model.__class__.__name__ == "DPTForDepthEstimation":
# Confirm out_indices propagated to backbone
self.assertEqual(len(model.backbone.out_indices), 2)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.use_pretrained_backbone = True
config.backbone_config = None
config.backbone_kwargs = {"out_indices": [-2, -1]}
# Force load_backbone path
config.is_hybrid = False
# Load a timm backbone
config.backbone = "resnet18"
config.use_timm_backbone = True
_validate_backbone_init()
# Load a HF backbone
config.backbone = "facebook/dinov2-small"
config.use_timm_backbone = False
_validate_backbone_init()
@slow
def test_model_from_pretrained(self):
model_name = "Intel/dpt-large"
model = DPTModel.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
@slow
class DPTModelIntegrationTest(unittest.TestCase):
def test_inference_depth_estimation(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large").to(torch_device)
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = outputs.predicted_depth
# verify the predicted depth
expected_shape = torch.Size((1, 384, 384))
self.assertEqual(predicted_depth.shape, expected_shape)
expectations = Expectations(
{
(None, None): [[6.3199, 6.3629, 6.4148], [6.3850, 6.3615, 6.4166], [6.3519, 6.3176, 6.3575]],
("cuda", 8): [[6.3199, 6.3629, 6.4148], [6.3850, 6.3615, 6.4166], [6.3519, 6.3176, 6.3575]],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(outputs.predicted_depth[0, :3, :3], expected_slice, rtol=2e-4, atol=2e-4)
def test_inference_semantic_segmentation(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large-ade")
model = DPTForSemanticSegmentation.from_pretrained("Intel/dpt-large-ade").to(torch_device)
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, 150, 480, 480))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor(
[[4.0480, 4.2420, 4.4360], [4.3124, 4.5693, 4.8261], [4.5768, 4.8965, 5.2163]]
).to(torch_device)
torch.testing.assert_close(outputs.logits[0, 0, :3, :3], expected_slice, rtol=1e-4, atol=1e-4)
def test_post_processing_semantic_segmentation(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large-ade")
model = DPTForSemanticSegmentation.from_pretrained("Intel/dpt-large-ade").to(torch_device)
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
outputs.logits = outputs.logits.detach().cpu()
segmentation = image_processor.post_process_semantic_segmentation(outputs=outputs, target_sizes=[(500, 300)])
expected_shape = torch.Size((500, 300))
self.assertEqual(segmentation[0].shape, expected_shape)
segmentation = image_processor.post_process_semantic_segmentation(outputs=outputs)
expected_shape = torch.Size((480, 480))
self.assertEqual(segmentation[0].shape, expected_shape)
def test_post_processing_depth_estimation(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large")
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt")
# forward pass
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = image_processor.post_process_depth_estimation(outputs=outputs)[0]["predicted_depth"]
expected_shape = torch.Size((384, 384))
self.assertTrue(predicted_depth.shape == expected_shape)
predicted_depth_l = image_processor.post_process_depth_estimation(outputs=outputs, target_sizes=[(500, 500)])
predicted_depth_l = predicted_depth_l[0]["predicted_depth"]
expected_shape = torch.Size((500, 500))
self.assertTrue(predicted_depth_l.shape == expected_shape)
output_enlarged = torch.nn.functional.interpolate(
predicted_depth.unsqueeze(0).unsqueeze(1), size=(500, 500), mode="bicubic", align_corners=False
).squeeze()
self.assertTrue(output_enlarged.shape == expected_shape)
torch.testing.assert_close(predicted_depth_l, output_enlarged, atol=1e-3, rtol=1e-3)
@pytest.mark.torch_export_test
def test_export(self):
for strict in [True, False]:
with self.subTest(strict=strict):
if not is_torch_greater_or_equal_than_2_4:
self.skipTest(reason="This test requires torch >= 2.4 to run.")
model = DPTForSemanticSegmentation.from_pretrained("Intel/dpt-large-ade").to(torch_device).eval()
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-large-ade")
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
exported_program = torch.export.export(
model,
args=(inputs["pixel_values"],),
strict=strict,
)
with torch.no_grad():
eager_outputs = model(**inputs)
exported_outputs = exported_program.module().forward(inputs["pixel_values"])
self.assertEqual(eager_outputs.logits.shape, exported_outputs.logits.shape)
torch.testing.assert_close(eager_outputs.logits, exported_outputs.logits, rtol=1e-4, atol=1e-4)

346
transformers/tests/models/dpt/test_modeling_dpt_auto_backbone.py Normal file
View File

@@ -0,0 +1,346 @@
# Copyright 2023 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 DPT model."""
import unittest
from transformers import Dinov2Config, DPTConfig
from transformers.file_utils import is_torch_available, is_vision_available
from transformers.testing_utils import Expectations, require_torch, require_vision, slow, torch_device
from transformers.utils.import_utils import get_torch_major_and_minor_version
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 transformers import DPTForDepthEstimation
from transformers.models.auto.modeling_auto import MODEL_MAPPING_NAMES
if is_vision_available():
from PIL import Image
from transformers import DPTImageProcessor
class DPTModelTester:
def __init__(
self,
parent,
batch_size=2,
num_channels=3,
image_size=32,
patch_size=16,
use_labels=True,
num_labels=3,
is_training=True,
hidden_size=4,
num_hidden_layers=2,
num_attention_heads=2,
intermediate_size=8,
out_features=["stage1", "stage2"],
apply_layernorm=False,
reshape_hidden_states=False,
neck_hidden_sizes=[2, 2],
fusion_hidden_size=6,
):
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.patch_size = patch_size
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.out_features = out_features
self.apply_layernorm = apply_layernorm
self.reshape_hidden_states = reshape_hidden_states
self.use_labels = use_labels
self.num_labels = num_labels
self.is_training = is_training
self.neck_hidden_sizes = neck_hidden_sizes
self.fusion_hidden_size = fusion_hidden_size
# DPT's sequence length
self.seq_length = (self.image_size // self.patch_size) ** 2 + 1
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.image_size, self.image_size], self.num_labels)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return DPTConfig(
backbone_config=self.get_backbone_config(),
backbone=None,
neck_hidden_sizes=self.neck_hidden_sizes,
fusion_hidden_size=self.fusion_hidden_size,
)
def get_backbone_config(self):
return Dinov2Config(
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,
is_training=self.is_training,
out_features=self.out_features,
reshape_hidden_states=self.reshape_hidden_states,
)
def create_and_check_for_depth_estimation(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = DPTForDepthEstimation(config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(result.predicted_depth.shape, (self.batch_size, self.image_size, self.image_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 DPTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as DPT does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (DPTForDepthEstimation,) if is_torch_available() else ()
pipeline_model_mapping = {"depth-estimation": DPTForDepthEstimation} if is_torch_available() else {}
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
test_torch_exportable = True
test_torch_exportable_strictly = get_torch_major_and_minor_version() != "2.7"
def setUp(self):
self.model_tester = DPTModelTester(self)
self.config_tester = ConfigTester(self, config_class=DPTConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="DPT with AutoBackbone does not have a base model and hence no input_embeddings")
def test_inputs_embeds(self):
pass
def test_for_depth_estimation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_depth_estimation(*config_and_inputs)
def test_training(self):
for model_class in self.all_model_classes:
if model_class.__name__ == "DPTForDepthEstimation":
continue
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
if model_class.__name__ in MODEL_MAPPING_NAMES.values():
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):
for model_class in self.all_model_classes:
if model_class.__name__ == "DPTForDepthEstimation":
continue
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.use_cache = False
config.return_dict = True
if model_class.__name__ in MODEL_MAPPING_NAMES.values() or not model_class.supports_gradient_checkpointing:
continue
model = model_class(config)
model.to(torch_device)
model.gradient_checkpointing_enable()
model.train()
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
loss = model(**inputs).loss
loss.backward()
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)
# Skip the check for the backbone
backbone_params = []
for name, module in model.named_modules():
if module.__class__.__name__ == "DPTViTHybridEmbeddings":
backbone_params = [f"{name}.{key}" for key in module.state_dict()]
break
for name, param in model.named_parameters():
if param.requires_grad:
if name in backbone_params:
continue
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
@unittest.skip(reason="DPT with AutoBackbone does not have a base model and hence no input_embeddings")
def test_model_get_set_embeddings(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
@slow
def test_model_from_pretrained(self):
model_name = "Intel/dpt-large"
model = DPTForDepthEstimation.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
@slow
class DPTModelIntegrationTest(unittest.TestCase):
def test_inference_depth_estimation_dinov2(self):
image_processor = DPTImageProcessor.from_pretrained("facebook/dpt-dinov2-small-kitti")
model = DPTForDepthEstimation.from_pretrained("facebook/dpt-dinov2-small-kitti").to(torch_device)
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = outputs.predicted_depth
# verify the predicted depth
expected_shape = torch.Size((1, 576, 736))
self.assertEqual(predicted_depth.shape, expected_shape)
expectations = Expectations(
{
(None, None): [[6.0336, 7.1502, 7.4130], [6.8977, 7.2383, 7.2268], [7.9180, 8.0525, 8.0134]],
("cuda", 8): [[6.0350, 7.1518, 7.4144], [6.8992, 7.2396, 7.2280], [7.9194, 8.0538, 8.0145]],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(outputs.predicted_depth[0, :3, :3], expected_slice, rtol=2e-4, atol=2e-4)
def test_inference_depth_estimation_beit(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-beit-base-384")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-beit-base-384").to(torch_device)
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = outputs.predicted_depth
# verify the predicted depth
expected_shape = torch.Size((1, 384, 384))
self.assertEqual(predicted_depth.shape, expected_shape)
expectations = Expectations(
{
(None, None): [
[2669.7061, 2663.7144, 2674.9399],
[2633.9326, 2650.9092, 2665.4270],
[2621.8271, 2632.0129, 2637.2290],
],
("cuda", 8): [
[2669.4292, 2663.4121, 2674.6233],
[2633.7400, 2650.7026, 2665.2085],
[2621.6572, 2631.8452, 2637.0525],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(outputs.predicted_depth[0, :3, :3], expected_slice, rtol=2e-4, atol=2e-4)
def test_inference_depth_estimation_swinv2(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-swinv2-tiny-256")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-swinv2-tiny-256").to(torch_device)
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = outputs.predicted_depth
# verify the predicted depth
expected_shape = torch.Size((1, 256, 256))
self.assertEqual(predicted_depth.shape, expected_shape)
expectations = Expectations(
{
(None, None): [
[1032.7719, 1025.1886, 1030.2661],
[1023.7619, 1021.0075, 1024.9121],
[1022.5667, 1018.8522, 1021.4145],
],
("cuda", 8): [
[1032.7170, 1025.0629, 1030.1941],
[1023.7309, 1020.9786, 1024.8594],
[1022.5233, 1018.8235, 1021.3312],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(outputs.predicted_depth[0, :3, :3], expected_slice, rtol=2e-4, atol=2e-4)

341
transformers/tests/models/dpt/test_modeling_dpt_hybrid.py Normal file
View File

@@ -0,0 +1,341 @@
# 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 DPT model."""
import unittest
from transformers import DPTConfig
from transformers.file_utils import is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
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 DPTForDepthEstimation, DPTForSemanticSegmentation, DPTModel
from transformers.models.auto.modeling_auto import MODEL_MAPPING_NAMES
if is_vision_available():
from PIL import Image
from transformers import DPTImageProcessor
class DPTModelTester:
def __init__(
self,
parent,
batch_size=2,
image_size=32,
patch_size=16,
num_channels=3,
is_training=True,
use_labels=True,
hidden_size=32,
num_hidden_layers=4,
backbone_out_indices=[0, 1, 2, 3],
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
initializer_range=0.02,
num_labels=3,
backbone_featmap_shape=[1, 32, 24, 24],
neck_hidden_sizes=[16, 16, 32, 32],
is_hybrid=True,
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.backbone_out_indices = backbone_out_indices
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.initializer_range = initializer_range
self.num_labels = num_labels
self.backbone_featmap_shape = backbone_featmap_shape
self.scope = scope
self.is_hybrid = is_hybrid
self.neck_hidden_sizes = neck_hidden_sizes
# sequence length of DPT = num_patches + 1 (we add 1 for the [CLS] token)
num_patches = (image_size // patch_size) ** 2
self.seq_length = num_patches + 1
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.image_size, self.image_size], self.num_labels)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
backbone_config = {
"global_padding": "same",
"layer_type": "bottleneck",
"depths": [3, 4, 9],
"out_features": ["stage1", "stage2", "stage3"],
"embedding_dynamic_padding": True,
"hidden_sizes": [16, 16, 32, 32],
"num_groups": 2,
}
return DPTConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
fusion_hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
backbone_out_indices=self.backbone_out_indices,
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,
is_hybrid=self.is_hybrid,
backbone_config=backbone_config,
backbone=None,
backbone_featmap_shape=self.backbone_featmap_shape,
neck_hidden_sizes=self.neck_hidden_sizes,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DPTModel(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_depth_estimation(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = DPTForDepthEstimation(config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(result.predicted_depth.shape, (self.batch_size, self.image_size, self.image_size))
def create_and_check_for_semantic_segmentation(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = DPTForSemanticSegmentation(config)
model.to(torch_device)
model.eval()
result = model(pixel_values, labels=labels)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_labels, self.image_size, self.image_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 DPTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as DPT does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (DPTModel, DPTForDepthEstimation, DPTForSemanticSegmentation) if is_torch_available() else ()
pipeline_model_mapping = (
{
"depth-estimation": DPTForDepthEstimation,
"feature-extraction": DPTModel,
"image-segmentation": DPTForSemanticSegmentation,
}
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 = DPTModelTester(self)
self.config_tester = ConfigTester(self, config_class=DPTConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_batching_equivalence(self, atol=2e-5, rtol=2e-5):
super().test_batching_equivalence(atol=atol, rtol=rtol)
@unittest.skip(reason="DPT 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_depth_estimation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_depth_estimation(*config_and_inputs)
def test_for_semantic_segmentation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_semantic_segmentation(*config_and_inputs)
def test_training(self):
for model_class in self.all_model_classes:
if model_class.__name__ == "DPTForDepthEstimation":
continue
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
if model_class.__name__ in MODEL_MAPPING_NAMES.values():
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):
for model_class in self.all_model_classes:
if model_class.__name__ == "DPTForDepthEstimation":
continue
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.use_cache = False
config.return_dict = True
if model_class.__name__ in MODEL_MAPPING_NAMES.values() or not model_class.supports_gradient_checkpointing:
continue
model = model_class(config)
model.to(torch_device)
model.gradient_checkpointing_enable()
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_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)
# Skip the check for the backbone
backbone_params = []
for name, module in model.named_modules():
if module.__class__.__name__ == "DPTViTHybridEmbeddings":
backbone_params = [f"{name}.{key}" for key in module.state_dict()]
break
for name, param in model.named_parameters():
if param.requires_grad:
if name in backbone_params:
continue
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
@slow
def test_model_from_pretrained(self):
model_name = "Intel/dpt-hybrid-midas"
model = DPTModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_raise_readout_type(self):
# We do this test only for DPTForDepthEstimation since it is the only model that uses readout_type
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
config.readout_type = "add"
with self.assertRaises(ValueError):
_ = DPTForDepthEstimation(config)
# 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
@slow
class DPTModelIntegrationTest(unittest.TestCase):
def test_inference_depth_estimation(self):
image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to(torch_device)
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = outputs.predicted_depth
# verify the predicted depth
expected_shape = torch.Size((1, 384, 384))
self.assertEqual(predicted_depth.shape, expected_shape)
expected_slice = torch.tensor(
[[[5.6437, 5.6146, 5.6511], [5.4371, 5.5649, 5.5958], [5.5215, 5.5184, 5.5293]]]
).to(torch_device)
torch.testing.assert_close(outputs.predicted_depth[:3, :3, :3] / 100, expected_slice, rtol=1e-4, atol=1e-4)