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

Add DETA (#20983)

Browse Source 
* First draft

* Add initial draft of conversion script

* Convert all weights

* Fix config

* Add image processor

* Fix DetaImageProcessor

* Run make fix copies

* Remove timm dependency

* Fix dummy objects

* Improve loss function

* Remove conv_encoder attribute

* Update conversion scripts

* Improve postprocessing + docs

* Fix copied from statements

* Add tests

* Improve postprocessing

* Improve postprocessing

* Update READMEs

* More improvements

* Fix rebase

* Add is_torchvision_available

* Add torchvision dependency

* Fix typo and README

* Fix bug

* Add copied from

* Fix style

* Apply suggestions

* Fix thanks to @ydshieh

* Fix another dependency check

* Simplify image processor

* Add scipy

* Improve code

* Add threshold argument

* Fix bug

* Set default threshold

* Improve integration test

* Add another integration test

* Update setup.py

* Address review

* Improve deformable attention function

* Improve copied from

* Use relative imports

* Address review

* Replace assertions

* Address review

* Update dummies

* Remove dummies

* Address comments, update READMEs

* Remove custom kernel code

* Add image processor tests

* Add requires_backends

* Add minor comment

* Update scripts

* Update organization name

* Fix defaults, add doc tests

* Add id2label for object 365

* Fix tests

* Update task guide
This commit is contained in:
NielsRogge
2023-01-31 10:43:10 +01:00
committed by GitHub
parent 98d88b23f5
commit 5451f8896c
38 changed files with 5939 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/models/deta/__init__.py Normal file
View File

356
tests/models/deta/test_image_processing_deta.py Normal file
View File

@@ -0,0 +1,356 @@
# 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 json
import pathlib
import unittest
import numpy as np
from transformers.testing_utils import require_torch, require_vision, slow
from transformers.utils import is_torch_available, is_vision_available
from ...test_image_processing_common import ImageProcessingSavingTestMixin, prepare_image_inputs
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import DetaImageProcessor
class DetaImageProcessingTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
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_rescale=True,
rescale_factor=1 / 255,
do_pad=True,
):
# by setting size["longest_edge"] > max_resolution we're effectively not testing this :p
size = size if size is not None else {"shortest_edge": 18, "longest_edge": 1333}
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
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_rescale = do_rescale
self.rescale_factor = rescale_factor
self.do_pad = do_pad
def prepare_image_processor_dict(self):
return {
"do_resize": self.do_resize,
"size": self.size,
"do_normalize": self.do_normalize,
"image_mean": self.image_mean,
"image_std": self.image_std,
"do_rescale": self.do_rescale,
"rescale_factor": self.rescale_factor,
"do_pad": self.do_pad,
}
def get_expected_values(self, image_inputs, batched=False):
"""
This function computes the expected height and width when providing images to DetaImageProcessor,
assuming do_resize is set to True with a scalar size.
"""
if not batched:
image = image_inputs[0]
if isinstance(image, Image.Image):
w, h = image.size
else:
h, w = image.shape[1], image.shape[2]
if w < h:
expected_height = int(self.size["shortest_edge"] * h / w)
expected_width = self.size["shortest_edge"]
elif w > h:
expected_height = self.size["shortest_edge"]
expected_width = int(self.size["shortest_edge"] * w / h)
else:
expected_height = self.size["shortest_edge"]
expected_width = self.size["shortest_edge"]
else:
expected_values = []
for image in image_inputs:
expected_height, expected_width = self.get_expected_values([image])
expected_values.append((expected_height, expected_width))
expected_height = max(expected_values, key=lambda item: item[0])[0]
expected_width = max(expected_values, key=lambda item: item[1])[1]
return expected_height, expected_width
@require_torch
@require_vision
class DetaImageProcessingTest(ImageProcessingSavingTestMixin, unittest.TestCase):
image_processing_class = DetaImageProcessor if is_vision_available() else None
def setUp(self):
self.image_processor_tester = DetaImageProcessingTester(self)
@property
def image_processor_dict(self):
return self.image_processor_tester.prepare_image_processor_dict()
def test_image_processor_properties(self):
image_processing = self.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, "do_rescale"))
self.assertTrue(hasattr(image_processing, "do_pad"))
self.assertTrue(hasattr(image_processing, "size"))
def test_image_processor_from_dict_with_kwargs(self):
image_processor = self.image_processing_class.from_dict(self.image_processor_dict)
self.assertEqual(image_processor.size, {"shortest_edge": 18, "longest_edge": 1333})
self.assertEqual(image_processor.do_pad, True)
def test_batch_feature(self):
pass
def test_call_pil(self):
# Initialize image_processing
image_processing = self.image_processing_class(**self.image_processor_dict)
# create random PIL images
image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False)
for image in image_inputs:
self.assertIsInstance(image, Image.Image)
# Test not batched input
encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs)
self.assertEqual(
encoded_images.shape,
(1, self.image_processor_tester.num_channels, expected_height, expected_width),
)
# Test batched
expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs, batched=True)
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
self.assertEqual(
encoded_images.shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.num_channels,
expected_height,
expected_width,
),
)
def test_call_numpy(self):
# Initialize image_processing
image_processing = self.image_processing_class(**self.image_processor_dict)
# create random numpy tensors
image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False, numpify=True)
for image in image_inputs:
self.assertIsInstance(image, np.ndarray)
# Test not batched input
encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs)
self.assertEqual(
encoded_images.shape,
(1, self.image_processor_tester.num_channels, expected_height, expected_width),
)
# Test batched
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs, batched=True)
self.assertEqual(
encoded_images.shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.num_channels,
expected_height,
expected_width,
),
)
def test_call_pytorch(self):
# Initialize image_processing
image_processing = self.image_processing_class(**self.image_processor_dict)
# create random PyTorch tensors
image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False, torchify=True)
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
# Test not batched input
encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs)
self.assertEqual(
encoded_images.shape,
(1, self.image_processor_tester.num_channels, expected_height, expected_width),
)
# Test batched
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
expected_height, expected_width = self.image_processor_tester.get_expected_values(image_inputs, batched=True)
self.assertEqual(
encoded_images.shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.num_channels,
expected_height,
expected_width,
),
)
def test_equivalence_pad_and_create_pixel_mask(self):
# Initialize image_processings
image_processing_1 = self.image_processing_class(**self.image_processor_dict)
image_processing_2 = self.image_processing_class(do_resize=False, do_normalize=False, do_rescale=False)
# create random PyTorch tensors
image_inputs = prepare_image_inputs(self.image_processor_tester, equal_resolution=False, torchify=True)
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
# Test whether the method "pad_and_return_pixel_mask" and calling the image processor return the same tensors
encoded_images_with_method = image_processing_1.pad_and_create_pixel_mask(image_inputs, return_tensors="pt")
encoded_images = image_processing_2(image_inputs, return_tensors="pt")
self.assertTrue(
torch.allclose(encoded_images_with_method["pixel_values"], encoded_images["pixel_values"], atol=1e-4)
)
self.assertTrue(
torch.allclose(encoded_images_with_method["pixel_mask"], encoded_images["pixel_mask"], atol=1e-4)
)
@slow
def test_call_pytorch_with_coco_detection_annotations(self):
# prepare image and target
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
with open("./tests/fixtures/tests_samples/COCO/coco_annotations.txt", "r") as f:
target = json.loads(f.read())
target = {"image_id": 39769, "annotations": target}
# encode them
image_processing = DetaImageProcessor()
encoding = image_processing(images=image, annotations=target, return_tensors="pt")
# verify pixel values
expected_shape = torch.Size([1, 3, 800, 1066])
self.assertEqual(encoding["pixel_values"].shape, expected_shape)
expected_slice = torch.tensor([0.2796, 0.3138, 0.3481])
self.assertTrue(torch.allclose(encoding["pixel_values"][0, 0, 0, :3], expected_slice, atol=1e-4))
# verify area
expected_area = torch.tensor([5887.9600, 11250.2061, 489353.8438, 837122.7500, 147967.5156, 165732.3438])
self.assertTrue(torch.allclose(encoding["labels"][0]["area"], expected_area))
# verify boxes
expected_boxes_shape = torch.Size([6, 4])
self.assertEqual(encoding["labels"][0]["boxes"].shape, expected_boxes_shape)
expected_boxes_slice = torch.tensor([0.5503, 0.2765, 0.0604, 0.2215])
self.assertTrue(torch.allclose(encoding["labels"][0]["boxes"][0], expected_boxes_slice, atol=1e-3))
# verify image_id
expected_image_id = torch.tensor([39769])
self.assertTrue(torch.allclose(encoding["labels"][0]["image_id"], expected_image_id))
# verify is_crowd
expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0])
self.assertTrue(torch.allclose(encoding["labels"][0]["iscrowd"], expected_is_crowd))
# verify class_labels
expected_class_labels = torch.tensor([75, 75, 63, 65, 17, 17])
self.assertTrue(torch.allclose(encoding["labels"][0]["class_labels"], expected_class_labels))
# verify orig_size
expected_orig_size = torch.tensor([480, 640])
self.assertTrue(torch.allclose(encoding["labels"][0]["orig_size"], expected_orig_size))
# verify size
expected_size = torch.tensor([800, 1066])
self.assertTrue(torch.allclose(encoding["labels"][0]["size"], expected_size))
@slow
def test_call_pytorch_with_coco_panoptic_annotations(self):
# prepare image, target and masks_path
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
with open("./tests/fixtures/tests_samples/COCO/coco_panoptic_annotations.txt", "r") as f:
target = json.loads(f.read())
target = {"file_name": "000000039769.png", "image_id": 39769, "segments_info": target}
masks_path = pathlib.Path("./tests/fixtures/tests_samples/COCO/coco_panoptic")
# encode them
image_processing = DetaImageProcessor(format="coco_panoptic")
encoding = image_processing(images=image, annotations=target, masks_path=masks_path, return_tensors="pt")
# verify pixel values
expected_shape = torch.Size([1, 3, 800, 1066])
self.assertEqual(encoding["pixel_values"].shape, expected_shape)
expected_slice = torch.tensor([0.2796, 0.3138, 0.3481])
self.assertTrue(torch.allclose(encoding["pixel_values"][0, 0, 0, :3], expected_slice, atol=1e-4))
# verify area
expected_area = torch.tensor([147979.6875, 165527.0469, 484638.5938, 11292.9375, 5879.6562, 7634.1147])
self.assertTrue(torch.allclose(encoding["labels"][0]["area"], expected_area))
# verify boxes
expected_boxes_shape = torch.Size([6, 4])
self.assertEqual(encoding["labels"][0]["boxes"].shape, expected_boxes_shape)
expected_boxes_slice = torch.tensor([0.2625, 0.5437, 0.4688, 0.8625])
self.assertTrue(torch.allclose(encoding["labels"][0]["boxes"][0], expected_boxes_slice, atol=1e-3))
# verify image_id
expected_image_id = torch.tensor([39769])
self.assertTrue(torch.allclose(encoding["labels"][0]["image_id"], expected_image_id))
# verify is_crowd
expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0])
self.assertTrue(torch.allclose(encoding["labels"][0]["iscrowd"], expected_is_crowd))
# verify class_labels
expected_class_labels = torch.tensor([17, 17, 63, 75, 75, 93])
self.assertTrue(torch.allclose(encoding["labels"][0]["class_labels"], expected_class_labels))
# verify masks
expected_masks_sum = 822873
self.assertEqual(encoding["labels"][0]["masks"].sum().item(), expected_masks_sum)
# verify orig_size
expected_orig_size = torch.tensor([480, 640])
self.assertTrue(torch.allclose(encoding["labels"][0]["orig_size"], expected_orig_size))
# verify size
expected_size = torch.tensor([800, 1066])
self.assertTrue(torch.allclose(encoding["labels"][0]["size"], expected_size))

522
tests/models/deta/test_modeling_deta.py Normal file
View File

@@ -0,0 +1,522 @@
# 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 DETA model. """
import inspect
import math
import unittest
from transformers import DetaConfig, is_torch_available, is_torchvision_available, is_vision_available
from transformers.file_utils import cached_property
from transformers.testing_utils import require_torchvision, require_vision, slow, torch_device
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
if is_torch_available():
import torch
if is_torchvision_available():
from transformers import DetaForObjectDetection, DetaModel
if is_vision_available():
from PIL import Image
from transformers import AutoImageProcessor
class DetaModelTester:
def __init__(
self,
parent,
batch_size=8,
is_training=True,
use_labels=True,
hidden_size=256,
num_hidden_layers=2,
num_attention_heads=8,
intermediate_size=4,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
num_queries=12,
num_channels=3,
image_size=196,
n_targets=8,
num_labels=91,
num_feature_levels=4,
encoder_n_points=2,
decoder_n_points=6,
two_stage=False,
):
self.parent = parent
self.batch_size = batch_size
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.num_queries = num_queries
self.num_channels = num_channels
self.image_size = image_size
self.n_targets = n_targets
self.num_labels = num_labels
self.num_feature_levels = num_feature_levels
self.encoder_n_points = encoder_n_points
self.decoder_n_points = decoder_n_points
self.two_stage = two_stage
# we also set the expected seq length for both encoder and decoder
self.encoder_seq_length = (
math.ceil(self.image_size / 8) ** 2
+ math.ceil(self.image_size / 16) ** 2
+ math.ceil(self.image_size / 32) ** 2
+ math.ceil(self.image_size / 64) ** 2
)
self.decoder_seq_length = self.num_queries
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
pixel_mask = torch.ones([self.batch_size, self.image_size, self.image_size], device=torch_device)
labels = None
if self.use_labels:
# labels is a list of Dict (each Dict being the labels for a given example in the batch)
labels = []
for i in range(self.batch_size):
target = {}
target["class_labels"] = torch.randint(
high=self.num_labels, size=(self.n_targets,), device=torch_device
)
target["boxes"] = torch.rand(self.n_targets, 4, device=torch_device)
target["masks"] = torch.rand(self.n_targets, self.image_size, self.image_size, device=torch_device)
labels.append(target)
config = self.get_config()
return config, pixel_values, pixel_mask, labels
def get_config(self):
return DetaConfig(
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
num_queries=self.num_queries,
num_labels=self.num_labels,
num_feature_levels=self.num_feature_levels,
encoder_n_points=self.encoder_n_points,
decoder_n_points=self.decoder_n_points,
two_stage=self.two_stage,
)
def prepare_config_and_inputs_for_common(self):
config, pixel_values, pixel_mask, labels = self.prepare_config_and_inputs()
inputs_dict = {"pixel_values": pixel_values, "pixel_mask": pixel_mask}
return config, inputs_dict
def create_and_check_deta_model(self, config, pixel_values, pixel_mask, labels):
model = DetaModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
result = model(pixel_values)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.num_queries, self.hidden_size))
def create_and_check_deta_object_detection_head_model(self, config, pixel_values, pixel_mask, labels):
model = DetaForObjectDetection(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels))
self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4))
result = model(pixel_values=pixel_values, pixel_mask=pixel_mask, labels=labels)
self.parent.assertEqual(result.loss.shape, ())
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels))
self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4))
@require_torchvision
class DetaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (DetaModel, DetaForObjectDetection) if is_torchvision_available() else ()
is_encoder_decoder = True
test_torchscript = False
test_pruning = False
test_head_masking = False
test_missing_keys = False
# special case for head models
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
if return_labels:
if model_class.__name__ == "DetaForObjectDetection":
labels = []
for i in range(self.model_tester.batch_size):
target = {}
target["class_labels"] = torch.ones(
size=(self.model_tester.n_targets,), device=torch_device, dtype=torch.long
)
target["boxes"] = torch.ones(
self.model_tester.n_targets, 4, device=torch_device, dtype=torch.float
)
target["masks"] = torch.ones(
self.model_tester.n_targets,
self.model_tester.image_size,
self.model_tester.image_size,
device=torch_device,
dtype=torch.float,
)
labels.append(target)
inputs_dict["labels"] = labels
return inputs_dict
def setUp(self):
self.model_tester = DetaModelTester(self)
self.config_tester = ConfigTester(self, config_class=DetaConfig, has_text_modality=False)
def test_config(self):
# we don't test common_properties and arguments_init as these don't apply for DETA
self.config_tester.create_and_test_config_to_json_string()
self.config_tester.create_and_test_config_to_json_file()
self.config_tester.create_and_test_config_from_and_save_pretrained()
self.config_tester.create_and_test_config_with_num_labels()
self.config_tester.check_config_can_be_init_without_params()
def test_deta_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_deta_model(*config_and_inputs)
def test_deta_object_detection_head_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_deta_object_detection_head_model(*config_and_inputs)
@unittest.skip(reason="DETA does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="DETA does not have a get_input_embeddings method")
def test_model_common_attributes(self):
pass
@unittest.skip(reason="DETA is not a generative model")
def test_generate_without_input_ids(self):
pass
@unittest.skip(reason="DETA does not use token embeddings")
def test_resize_tokens_embeddings(self):
pass
@unittest.skip(reason="Feed forward chunking is not implemented")
def test_feed_forward_chunking(self):
pass
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
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))
attentions = outputs.encoder_attentions
self.assertEqual(len(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.encoder_attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads,
self.model_tester.num_feature_levels,
self.model_tester.encoder_n_points,
],
)
out_len = len(outputs)
correct_outlen = 8
# loss is at first position
if "labels" in inputs_dict:
correct_outlen += 1 # loss is added to beginning
# Object Detection model returns pred_logits and pred_boxes
if model_class.__name__ == "DetaForObjectDetection":
correct_outlen += 2
self.assertEqual(out_len, correct_outlen)
# decoder attentions
decoder_attentions = outputs.decoder_attentions
self.assertIsInstance(decoder_attentions, (list, tuple))
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, self.model_tester.num_queries, self.model_tester.num_queries],
)
# cross attentions
cross_attentions = outputs.cross_attentions
self.assertIsInstance(cross_attentions, (list, tuple))
self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(cross_attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads,
self.model_tester.num_feature_levels,
self.model_tester.decoder_n_points,
],
)
# 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))
if hasattr(self.model_tester, "num_hidden_states_types"):
added_hidden_states = self.model_tester.num_hidden_states_types
elif self.is_encoder_decoder:
added_hidden_states = 2
else:
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.encoder_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,
self.model_tester.num_feature_levels,
self.model_tester.encoder_n_points,
],
)
# removed retain_grad and grad on decoder_hidden_states, as queries don't require grad
def test_retain_grad_hidden_states_attentions(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True
config.output_attentions = True
# no need to test all models as different heads yield the same functionality
model_class = self.all_model_classes[0]
model = model_class(config)
model.to(torch_device)
inputs = self._prepare_for_class(inputs_dict, model_class)
outputs = model(**inputs)
# we take the second output since last_hidden_state is the second item
output = outputs[1]
encoder_hidden_states = outputs.encoder_hidden_states[0]
encoder_attentions = outputs.encoder_attentions[0]
encoder_hidden_states.retain_grad()
encoder_attentions.retain_grad()
decoder_attentions = outputs.decoder_attentions[0]
decoder_attentions.retain_grad()
cross_attentions = outputs.cross_attentions[0]
cross_attentions.retain_grad()
output.flatten()[0].backward(retain_graph=True)
self.assertIsNotNone(encoder_hidden_states.grad)
self.assertIsNotNone(encoder_attentions.grad)
self.assertIsNotNone(decoder_attentions.grad)
self.assertIsNotNone(cross_attentions.grad)
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()]
if model.config.is_encoder_decoder:
expected_arg_names = ["pixel_values", "pixel_mask"]
expected_arg_names.extend(
["head_mask", "decoder_head_mask", "encoder_outputs"]
if "head_mask" and "decoder_head_mask" in arg_names
else []
)
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
else:
expected_arg_names = ["pixel_values", "pixel_mask"]
self.assertListEqual(arg_names[:1], expected_arg_names)
@unittest.skip(reason="Model doesn't use tied weights")
def test_tied_model_weights_key_ignore(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)
for name, param in model.named_parameters():
if param.requires_grad:
if param.requires_grad:
if (
"level_embed" in name
or "sampling_offsets.bias" in name
or "value_proj" in name
or "output_proj" in name
or "reference_points" in name
):
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",
)
TOLERANCE = 1e-4
# 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_torchvision
@require_vision
@slow
class DetaModelIntegrationTests(unittest.TestCase):
@cached_property
def default_image_processor(self):
return AutoImageProcessor.from_pretrained("jozhang97/deta-resnet-50") if is_vision_available() else None
def test_inference_object_detection_head(self):
model = DetaForObjectDetection.from_pretrained("jozhang97/deta-resnet-50").to(torch_device)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
with torch.no_grad():
outputs = model(**inputs)
expected_shape_logits = torch.Size((1, 300, model.config.num_labels))
self.assertEqual(outputs.logits.shape, expected_shape_logits)
expected_logits = torch.tensor(
[[-7.3978, -2.5406, -4.1668], [-8.2684, -3.9933, -3.8096], [-7.0515, -3.7973, -5.8516]]
).to(torch_device)
expected_boxes = torch.tensor(
[[0.5043, 0.4973, 0.9998], [0.2542, 0.5489, 0.4748], [0.5490, 0.2765, 0.0570]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_logits, atol=1e-4))
expected_shape_boxes = torch.Size((1, 300, 4))
self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4))
# verify postprocessing
results = image_processor.post_process_object_detection(
outputs, threshold=0.3, target_sizes=[image.size[::-1]]
)[0]
expected_scores = torch.tensor([0.6392, 0.6276, 0.5546, 0.5260, 0.4706], device=torch_device)
expected_labels = [75, 17, 17, 75, 63]
expected_slice_boxes = torch.tensor([40.5866, 73.2107, 176.1421, 117.1751], device=torch_device)
self.assertTrue(torch.allclose(results["scores"], expected_scores, atol=1e-4))
self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
self.assertTrue(torch.allclose(results["boxes"][0, :], expected_slice_boxes))
def test_inference_object_detection_head_swin_backbone(self):
model = DetaForObjectDetection.from_pretrained("jozhang97/deta-swin-large").to(torch_device)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
with torch.no_grad():
outputs = model(**inputs)
expected_shape_logits = torch.Size((1, 300, model.config.num_labels))
self.assertEqual(outputs.logits.shape, expected_shape_logits)
expected_logits = torch.tensor(
[[-7.6308, -2.8485, -5.3737], [-7.2037, -4.5505, -4.8027], [-7.2943, -4.2611, -4.6617]]
).to(torch_device)
expected_boxes = torch.tensor(
[[0.4987, 0.4969, 0.9999], [0.2549, 0.5498, 0.4805], [0.5498, 0.2757, 0.0569]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_logits, atol=1e-4))
expected_shape_boxes = torch.Size((1, 300, 4))
self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4))
# verify postprocessing
results = image_processor.post_process_object_detection(
outputs, threshold=0.3, target_sizes=[image.size[::-1]]
)[0]
expected_scores = torch.tensor([0.6831, 0.6826, 0.5684, 0.5464, 0.4392], device=torch_device)
expected_labels = [17, 17, 75, 75, 63]
expected_slice_boxes = torch.tensor([345.8478, 23.6754, 639.8562, 372.8265], device=torch_device)
self.assertTrue(torch.allclose(results["scores"], expected_scores, atol=1e-4))
self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
self.assertTrue(torch.allclose(results["boxes"][0, :], expected_slice_boxes))