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

Add DPT (#15991)

Browse Source 
* First draft

* More improvements

* Add fusion blocks

* Make conversion script work for dpt_large

* Make conversion script work

* Improve implementation

* Improve conversion script

* Add DPTForSemanticSegmentation

* Make conversion work for semantic segmentation

* Add tests

* Remove print statements

* First draft

* Redesign neck

* Improve tests

* Improve implementation some more

* Make neck output list of tensors

* Improve neck and feature extractor

* Fix integration tests

* Make more tests pass

* Make all tests pass

* Add missing config archive map

* Add in_index attribute to make heads accept list of tensors

* Apply suggestions from code review

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Apply some more suggestions

* Add copied from statements

* Remove assert

* Apply suggestions from code review

* Apply suggestions from code review

* Remove DPTInterpolate in favor of nn.Upsample

* Add comments

* Apply suggestions from code review

* Apply suggestions from code review

* Add proposed design

* Update design

* Add DPTReassembleLayer

* Add DPTFeatureFusionStage

* Apply more suggestions from code review

* Apply suggestions from code review

* Apply suggestions from code review

* Fix rebase

* Update in_index and out_indices

* Fix conversion script

* Fix code quality

* Add model to toctree and use DepthEstimatorOutput

* Fix rebase

* Fix code examples

* Improve code

* Fix copied from statements

* Apply suggestions from code review

* Remove compute_loss method

* Apply suggestions from code review

* Fix documentation tests file

* Remove test.py file

* Improve doc example

Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Niels Rogge <nielsrogge@nielss-mbp.home>
This commit is contained in:
NielsRogge
2022-03-28 16:28:10 +02:00
committed by GitHub
parent 7ca4633555
commit 979b039c89
24 changed files with 2565 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/dpt/__init__.py Normal file
View File

188
tests/dpt/test_feature_extraction_dpt.py Normal file
View File

@@ -0,0 +1,188 @@
# coding=utf-8
# 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 transformers.file_utils import is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_vision
from ..test_feature_extraction_common import FeatureExtractionSavingTestMixin, prepare_image_inputs
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import DPTFeatureExtractor
class DPTFeatureExtractionTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=400,
do_resize=True,
size=18,
do_normalize=True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
):
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
def prepare_feat_extract_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,
}
@require_torch
@require_vision
class DPTFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
feature_extraction_class = DPTFeatureExtractor if is_vision_available() else None
def setUp(self):
self.feature_extract_tester = DPTFeatureExtractionTester(self)
@property
def feat_extract_dict(self):
return self.feature_extract_tester.prepare_feat_extract_dict()
def test_feat_extract_properties(self):
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
self.assertTrue(hasattr(feature_extractor, "image_mean"))
self.assertTrue(hasattr(feature_extractor, "image_std"))
self.assertTrue(hasattr(feature_extractor, "do_normalize"))
self.assertTrue(hasattr(feature_extractor, "do_resize"))
self.assertTrue(hasattr(feature_extractor, "size"))
def test_call_pil(self):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
# create random PIL images
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False)
for image in image_inputs:
self.assertIsInstance(image, Image.Image)
# Test not batched input
encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
1,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
),
)
# Test batched
encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.feature_extract_tester.batch_size,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
),
)
def test_call_numpy(self):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
# create random numpy tensors
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, numpify=True)
for image in image_inputs:
self.assertIsInstance(image, np.ndarray)
# Test not batched input
encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
1,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
),
)
# Test batched
encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.feature_extract_tester.batch_size,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
),
)
def test_call_pytorch(self):
# Initialize feature_extractor
feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
# create random PyTorch tensors
image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, torchify=True)
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
# Test not batched input
encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
1,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
),
)
# Test batched
encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.feature_extract_tester.batch_size,
self.feature_extract_tester.num_channels,
self.feature_extract_tester.size,
self.feature_extract_tester.size,
),
)

393
tests/dpt/test_modeling_dpt.py Normal file
View File

@@ -0,0 +1,393 @@
# coding=utf-8
# 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 inspect
import unittest
from transformers import DPTConfig
from transformers.file_utils import is_torch_available, is_vision_available
from transformers.models.auto import get_values
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from ..test_configuration_common import ConfigTester
from ..test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
if is_torch_available():
import torch
from torch import nn
from transformers import MODEL_MAPPING, DPTForDepthEstimation, DPTForSemanticSegmentation, DPTModel
from transformers.models.dpt.modeling_dpt import DPT_PRETRAINED_MODEL_ARCHIVE_LIST
if is_vision_available():
from PIL import Image
from transformers import DPTFeatureExtractor
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,
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
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,
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,
)
def create_and_check_model(self, config, pixel_values, labels):
model = DPTModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
num_patches = (config.image_size // config.patch_size) ** 2
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, 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, 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 ()
test_pruning = False
test_torchscript = False
test_resize_embeddings = False
test_head_masking = False
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_inputs_embeds(self):
# DPT does not use inputs_embeds
pass
def test_model_common_attributes(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_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["pixel_values"]
self.assertListEqual(arg_names[:1], expected_arg_names)
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_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
# in DPT, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
num_patches = (config.image_size // config.patch_size) ** 2
seq_len = num_patches + 1
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(len(outputs.attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_len, seq_len],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(out_len + 1, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_len, seq_len],
)
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.hidden_states
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
)
self.assertEqual(len(hidden_states), expected_num_layers)
# DPT has a different seq_length
num_patches = (config.image_size // config.patch_size) ** 2
seq_len = num_patches + 1
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_len, self.model_tester.hidden_size],
)
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)
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 in get_values(MODEL_MAPPING):
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 in get_values(MODEL_MAPPING) 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()
@slow
def test_model_from_pretrained(self):
for model_name in DPT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
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):
feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large")
model = DPTForDepthEstimation.from_pretrained("Intel/dpt-large").to(torch_device)
image = prepare_img()
inputs = feature_extractor(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(
[[6.3199, 6.3629, 6.4148], [6.3850, 6.3615, 6.4166], [6.3519, 6.3176, 6.3575]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.predicted_depth[0, :3, :3], expected_slice, atol=1e-4))
def test_inference_semantic_segmentation(self):
feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-large-ade")
model = DPTForSemanticSegmentation.from_pretrained("Intel/dpt-large-ade").to(torch_device)
image = prepare_img()
inputs = feature_extractor(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)
self.assertTrue(torch.allclose(outputs.logits[0, 0, :3, :3], expected_slice, atol=1e-4))

6
tests/test_modeling_common.py
View File

@@ -79,6 +79,7 @@ if is_torch_available():
MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
MODEL_FOR_QUESTION_ANSWERING_MAPPING,
MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING,
MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
@@ -177,6 +178,11 @@ class ModelTesterMixin:
inputs_dict["bool_masked_pos"] = torch.zeros(
(self.model_tester.batch_size, num_patches**2), dtype=torch.long, device=torch_device
)
elif model_class in get_values(MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING):
batch_size, num_channels, height, width = inputs_dict["pixel_values"].shape
inputs_dict["labels"] = torch.zeros(
[self.model_tester.batch_size, height, width], device=torch_device
).long()
return inputs_dict