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/conditional_detr/__init__.py Normal file
View File

607
transformers/tests/models/conditional_detr/test_image_processing_conditional_detr.py Normal file
View File

@@ -0,0 +1,607 @@
# 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_torchvision_available, is_vision_available
from ...test_image_processing_common import AnnotationFormatTestMixin, ImageProcessingTestMixin, prepare_image_inputs
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import ConditionalDetrImageProcessor
if is_torchvision_available():
from transformers import ConditionalDetrImageProcessorFast
class ConditionalDetrImageProcessingTester:
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 ConditionalDetrImageProcessor,
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
elif isinstance(image, np.ndarray):
h, w = image.shape[0], image.shape[1]
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
def expected_output_image_shape(self, images):
height, width = self.get_expected_values(images, batched=True)
return self.num_channels, height, width
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
return prepare_image_inputs(
batch_size=self.batch_size,
num_channels=self.num_channels,
min_resolution=self.min_resolution,
max_resolution=self.max_resolution,
equal_resolution=equal_resolution,
numpify=numpify,
torchify=torchify,
)
@require_torch
@require_vision
class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = ConditionalDetrImageProcessor if is_vision_available() else None
fast_image_processing_class = ConditionalDetrImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = ConditionalDetrImageProcessingTester(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"))
def test_image_processor_from_dict_with_kwargs(self):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class.from_dict(self.image_processor_dict)
self.assertEqual(image_processor.size, {"shortest_edge": 18, "longest_edge": 1333})
self.assertEqual(image_processor.do_pad, True)
image_processor = image_processing_class.from_dict(
self.image_processor_dict, size=42, max_size=84, pad_and_return_pixel_mask=False
)
self.assertEqual(image_processor.size, {"shortest_edge": 42, "longest_edge": 84})
self.assertEqual(image_processor.do_pad, False)
@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") as f:
target = json.loads(f.read())
target = {"image_id": 39769, "annotations": target}
for image_processing_class in self.image_processor_list:
# encode them
image_processing = image_processing_class.from_pretrained("microsoft/conditional-detr-resnet-50")
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])
torch.testing.assert_close(encoding["pixel_values"][0, 0, 0, :3], expected_slice, rtol=1e-4, atol=1e-4)
# verify area
expected_area = torch.tensor([5887.9600, 11250.2061, 489353.8438, 837122.7500, 147967.5156, 165732.3438])
torch.testing.assert_close(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])
torch.testing.assert_close(encoding["labels"][0]["boxes"][0], expected_boxes_slice, rtol=1e-3, atol=1e-3)
# verify image_id
expected_image_id = torch.tensor([39769])
torch.testing.assert_close(encoding["labels"][0]["image_id"], expected_image_id)
# verify is_crowd
expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0])
torch.testing.assert_close(encoding["labels"][0]["iscrowd"], expected_is_crowd)
# verify class_labels
expected_class_labels = torch.tensor([75, 75, 63, 65, 17, 17])
torch.testing.assert_close(encoding["labels"][0]["class_labels"], expected_class_labels)
# verify orig_size
expected_orig_size = torch.tensor([480, 640])
torch.testing.assert_close(encoding["labels"][0]["orig_size"], expected_orig_size)
# verify size
expected_size = torch.tensor([800, 1066])
torch.testing.assert_close(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") 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")
for image_processing_class in self.image_processor_list:
# encode them
image_processing = image_processing_class(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])
torch.testing.assert_close(encoding["pixel_values"][0, 0, 0, :3], expected_slice, rtol=1e-4, atol=1e-4)
# verify area
expected_area = torch.tensor([147979.6875, 165527.0469, 484638.5938, 11292.9375, 5879.6562, 7634.1147])
torch.testing.assert_close(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])
torch.testing.assert_close(encoding["labels"][0]["boxes"][0], expected_boxes_slice, rtol=1e-3, atol=1e-3)
# verify image_id
expected_image_id = torch.tensor([39769])
torch.testing.assert_close(encoding["labels"][0]["image_id"], expected_image_id)
# verify is_crowd
expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0])
torch.testing.assert_close(encoding["labels"][0]["iscrowd"], expected_is_crowd)
# verify class_labels
expected_class_labels = torch.tensor([17, 17, 63, 75, 75, 93])
torch.testing.assert_close(encoding["labels"][0]["class_labels"], expected_class_labels)
# verify masks
expected_masks_sum = 822873
relative_error = torch.abs(encoding["labels"][0]["masks"].sum() - expected_masks_sum) / expected_masks_sum
self.assertTrue(relative_error < 1e-3)
# verify orig_size
expected_orig_size = torch.tensor([480, 640])
torch.testing.assert_close(encoding["labels"][0]["orig_size"], expected_orig_size)
# verify size
expected_size = torch.tensor([800, 1066])
torch.testing.assert_close(encoding["labels"][0]["size"], expected_size)
@slow
# Copied from tests.models.detr.test_image_processing_detr.DetrImageProcessingTest.test_batched_coco_detection_annotations with Detr->ConditionalDetr, facebook/detr-resnet-50 ->microsoft/conditional-detr-resnet-50
def test_batched_coco_detection_annotations(self):
image_0 = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
image_1 = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png").resize((800, 800))
with open("./tests/fixtures/tests_samples/COCO/coco_annotations.txt") as f:
target = json.loads(f.read())
annotations_0 = {"image_id": 39769, "annotations": target}
annotations_1 = {"image_id": 39769, "annotations": target}
# Adjust the bounding boxes for the resized image
w_0, h_0 = image_0.size
w_1, h_1 = image_1.size
for i in range(len(annotations_1["annotations"])):
coords = annotations_1["annotations"][i]["bbox"]
new_bbox = [
coords[0] * w_1 / w_0,
coords[1] * h_1 / h_0,
coords[2] * w_1 / w_0,
coords[3] * h_1 / h_0,
]
annotations_1["annotations"][i]["bbox"] = new_bbox
images = [image_0, image_1]
annotations = [annotations_0, annotations_1]
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class()
encoding = image_processing(
images=images,
annotations=annotations,
return_segmentation_masks=True,
return_tensors="pt", # do_convert_annotations=True
)
# Check the pixel values have been padded
postprocessed_height, postprocessed_width = 800, 1066
expected_shape = torch.Size([2, 3, postprocessed_height, postprocessed_width])
self.assertEqual(encoding["pixel_values"].shape, expected_shape)
# Check the bounding boxes have been adjusted for padded images
self.assertEqual(encoding["labels"][0]["boxes"].shape, torch.Size([6, 4]))
self.assertEqual(encoding["labels"][1]["boxes"].shape, torch.Size([6, 4]))
expected_boxes_0 = torch.tensor(
[
[0.6879, 0.4609, 0.0755, 0.3691],
[0.2118, 0.3359, 0.2601, 0.1566],
[0.5011, 0.5000, 0.9979, 1.0000],
[0.5010, 0.5020, 0.9979, 0.9959],
[0.3284, 0.5944, 0.5884, 0.8112],
[0.8394, 0.5445, 0.3213, 0.9110],
]
)
expected_boxes_1 = torch.tensor(
[
[0.4130, 0.2765, 0.0453, 0.2215],
[0.1272, 0.2016, 0.1561, 0.0940],
[0.3757, 0.4933, 0.7488, 0.9865],
[0.3759, 0.5002, 0.7492, 0.9955],
[0.1971, 0.5456, 0.3532, 0.8646],
[0.5790, 0.4115, 0.3430, 0.7161],
]
)
torch.testing.assert_close(encoding["labels"][0]["boxes"], expected_boxes_0, atol=1e-3, rtol=1e-3)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, atol=1e-3, rtol=1e-3)
# Check the masks have also been padded
self.assertEqual(encoding["labels"][0]["masks"].shape, torch.Size([6, 800, 1066]))
self.assertEqual(encoding["labels"][1]["masks"].shape, torch.Size([6, 800, 1066]))
# Check if do_convert_annotations=False, then the annotations are not converted to centre_x, centre_y, width, height
# format and not in the range [0, 1]
encoding = image_processing(
images=images,
annotations=annotations,
return_segmentation_masks=True,
do_convert_annotations=False,
return_tensors="pt",
)
self.assertEqual(encoding["labels"][0]["boxes"].shape, torch.Size([6, 4]))
self.assertEqual(encoding["labels"][1]["boxes"].shape, torch.Size([6, 4]))
# Convert to absolute coordinates
unnormalized_boxes_0 = torch.vstack(
[
expected_boxes_0[:, 0] * postprocessed_width,
expected_boxes_0[:, 1] * postprocessed_height,
expected_boxes_0[:, 2] * postprocessed_width,
expected_boxes_0[:, 3] * postprocessed_height,
]
).T
unnormalized_boxes_1 = torch.vstack(
[
expected_boxes_1[:, 0] * postprocessed_width,
expected_boxes_1[:, 1] * postprocessed_height,
expected_boxes_1[:, 2] * postprocessed_width,
expected_boxes_1[:, 3] * postprocessed_height,
]
).T
# Convert from centre_x, centre_y, width, height to x_min, y_min, x_max, y_max
expected_boxes_0 = torch.vstack(
[
unnormalized_boxes_0[:, 0] - unnormalized_boxes_0[:, 2] / 2,
unnormalized_boxes_0[:, 1] - unnormalized_boxes_0[:, 3] / 2,
unnormalized_boxes_0[:, 0] + unnormalized_boxes_0[:, 2] / 2,
unnormalized_boxes_0[:, 1] + unnormalized_boxes_0[:, 3] / 2,
]
).T
expected_boxes_1 = torch.vstack(
[
unnormalized_boxes_1[:, 0] - unnormalized_boxes_1[:, 2] / 2,
unnormalized_boxes_1[:, 1] - unnormalized_boxes_1[:, 3] / 2,
unnormalized_boxes_1[:, 0] + unnormalized_boxes_1[:, 2] / 2,
unnormalized_boxes_1[:, 1] + unnormalized_boxes_1[:, 3] / 2,
]
).T
torch.testing.assert_close(encoding["labels"][0]["boxes"], expected_boxes_0, atol=1, rtol=1)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, atol=1, rtol=1)
# Copied from tests.models.detr.test_image_processing_detr.DetrImageProcessingTest.test_batched_coco_panoptic_annotations with Detr->ConditionalDetr
def test_batched_coco_panoptic_annotations(self):
# prepare image, target and masks_path
image_0 = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
image_1 = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png").resize((800, 800))
with open("./tests/fixtures/tests_samples/COCO/coco_panoptic_annotations.txt") as f:
target = json.loads(f.read())
annotation_0 = {"file_name": "000000039769.png", "image_id": 39769, "segments_info": target}
annotation_1 = {"file_name": "000000039769.png", "image_id": 39769, "segments_info": target}
w_0, h_0 = image_0.size
w_1, h_1 = image_1.size
for i in range(len(annotation_1["segments_info"])):
coords = annotation_1["segments_info"][i]["bbox"]
new_bbox = [
coords[0] * w_1 / w_0,
coords[1] * h_1 / h_0,
coords[2] * w_1 / w_0,
coords[3] * h_1 / h_0,
]
annotation_1["segments_info"][i]["bbox"] = new_bbox
masks_path = pathlib.Path("./tests/fixtures/tests_samples/COCO/coco_panoptic")
images = [image_0, image_1]
annotations = [annotation_0, annotation_1]
for image_processing_class in self.image_processor_list:
# encode them
image_processing = image_processing_class(format="coco_panoptic")
encoding = image_processing(
images=images,
annotations=annotations,
masks_path=masks_path,
return_tensors="pt",
return_segmentation_masks=True,
)
# Check the pixel values have been padded
postprocessed_height, postprocessed_width = 800, 1066
expected_shape = torch.Size([2, 3, postprocessed_height, postprocessed_width])
self.assertEqual(encoding["pixel_values"].shape, expected_shape)
# Check the bounding boxes have been adjusted for padded images
self.assertEqual(encoding["labels"][0]["boxes"].shape, torch.Size([6, 4]))
self.assertEqual(encoding["labels"][1]["boxes"].shape, torch.Size([6, 4]))
expected_boxes_0 = torch.tensor(
[
[0.2625, 0.5437, 0.4688, 0.8625],
[0.7719, 0.4104, 0.4531, 0.7125],
[0.5000, 0.4927, 0.9969, 0.9854],
[0.1688, 0.2000, 0.2063, 0.0917],
[0.5492, 0.2760, 0.0578, 0.2187],
[0.4992, 0.4990, 0.9984, 0.9979],
]
)
expected_boxes_1 = torch.tensor(
[
[0.1576, 0.3262, 0.2814, 0.5175],
[0.4634, 0.2463, 0.2720, 0.4275],
[0.3002, 0.2956, 0.5985, 0.5913],
[0.1013, 0.1200, 0.1238, 0.0550],
[0.3297, 0.1656, 0.0347, 0.1312],
[0.2997, 0.2994, 0.5994, 0.5987],
]
)
torch.testing.assert_close(encoding["labels"][0]["boxes"], expected_boxes_0, atol=1e-3, rtol=1e-3)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, atol=1e-3, rtol=1e-3)
# Check the masks have also been padded
self.assertEqual(encoding["labels"][0]["masks"].shape, torch.Size([6, 800, 1066]))
self.assertEqual(encoding["labels"][1]["masks"].shape, torch.Size([6, 800, 1066]))
# Check if do_convert_annotations=False, then the annotations are not converted to centre_x, centre_y, width, height
# format and not in the range [0, 1]
encoding = image_processing(
images=images,
annotations=annotations,
masks_path=masks_path,
return_segmentation_masks=True,
do_convert_annotations=False,
return_tensors="pt",
)
self.assertEqual(encoding["labels"][0]["boxes"].shape, torch.Size([6, 4]))
self.assertEqual(encoding["labels"][1]["boxes"].shape, torch.Size([6, 4]))
# Convert to absolute coordinates
unnormalized_boxes_0 = torch.vstack(
[
expected_boxes_0[:, 0] * postprocessed_width,
expected_boxes_0[:, 1] * postprocessed_height,
expected_boxes_0[:, 2] * postprocessed_width,
expected_boxes_0[:, 3] * postprocessed_height,
]
).T
unnormalized_boxes_1 = torch.vstack(
[
expected_boxes_1[:, 0] * postprocessed_width,
expected_boxes_1[:, 1] * postprocessed_height,
expected_boxes_1[:, 2] * postprocessed_width,
expected_boxes_1[:, 3] * postprocessed_height,
]
).T
# Convert from centre_x, centre_y, width, height to x_min, y_min, x_max, y_max
expected_boxes_0 = torch.vstack(
[
unnormalized_boxes_0[:, 0] - unnormalized_boxes_0[:, 2] / 2,
unnormalized_boxes_0[:, 1] - unnormalized_boxes_0[:, 3] / 2,
unnormalized_boxes_0[:, 0] + unnormalized_boxes_0[:, 2] / 2,
unnormalized_boxes_0[:, 1] + unnormalized_boxes_0[:, 3] / 2,
]
).T
expected_boxes_1 = torch.vstack(
[
unnormalized_boxes_1[:, 0] - unnormalized_boxes_1[:, 2] / 2,
unnormalized_boxes_1[:, 1] - unnormalized_boxes_1[:, 3] / 2,
unnormalized_boxes_1[:, 0] + unnormalized_boxes_1[:, 2] / 2,
unnormalized_boxes_1[:, 1] + unnormalized_boxes_1[:, 3] / 2,
]
).T
torch.testing.assert_close(encoding["labels"][0]["boxes"], expected_boxes_0, atol=1, rtol=1)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, atol=1, rtol=1)
# Copied from tests.models.detr.test_image_processing_detr.DetrImageProcessingTest.test_max_width_max_height_resizing_and_pad_strategy with Detr->ConditionalDetr
def test_max_width_max_height_resizing_and_pad_strategy(self):
for image_processing_class in self.image_processor_list:
image_1 = torch.ones([200, 100, 3], dtype=torch.uint8)
# do_pad=False, max_height=100, max_width=100, image=200x100 -> 100x50
image_processor = image_processing_class(
size={"max_height": 100, "max_width": 100},
do_pad=False,
)
inputs = image_processor(images=[image_1], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 100, 50]))
# do_pad=False, max_height=300, max_width=100, image=200x100 -> 200x100
image_processor = image_processing_class(
size={"max_height": 300, "max_width": 100},
do_pad=False,
)
inputs = image_processor(images=[image_1], return_tensors="pt")
# do_pad=True, max_height=100, max_width=100, image=200x100 -> 100x100
image_processor = image_processing_class(
size={"max_height": 100, "max_width": 100}, do_pad=True, pad_size={"height": 100, "width": 100}
)
inputs = image_processor(images=[image_1], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 100, 100]))
# do_pad=True, max_height=300, max_width=100, image=200x100 -> 300x100
image_processor = image_processing_class(
size={"max_height": 300, "max_width": 100},
do_pad=True,
pad_size={"height": 301, "width": 101},
)
inputs = image_processor(images=[image_1], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 301, 101]))
### Check for batch
image_2 = torch.ones([100, 150, 3], dtype=torch.uint8)
# do_pad=True, max_height=150, max_width=100, images=[200x100, 100x150] -> 150x100
image_processor = image_processing_class(
size={"max_height": 150, "max_width": 100},
do_pad=True,
pad_size={"height": 150, "width": 100},
)
inputs = image_processor(images=[image_1, image_2], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([2, 3, 150, 100]))
def test_longest_edge_shortest_edge_resizing_strategy(self):
image_1 = torch.ones([958, 653, 3], dtype=torch.uint8)
# max size is set; width < height;
# do_pad=False, longest_edge=640, shortest_edge=640, image=958x653 -> 640x436
image_processor = ConditionalDetrImageProcessor(
size={"longest_edge": 640, "shortest_edge": 640},
do_pad=False,
)
inputs = image_processor(images=[image_1], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 640, 436]))
image_2 = torch.ones([653, 958, 3], dtype=torch.uint8)
# max size is set; height < width;
# do_pad=False, longest_edge=640, shortest_edge=640, image=653x958 -> 436x640
image_processor = ConditionalDetrImageProcessor(
size={"longest_edge": 640, "shortest_edge": 640},
do_pad=False,
)
inputs = image_processor(images=[image_2], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 436, 640]))
image_3 = torch.ones([100, 120, 3], dtype=torch.uint8)
# max size is set; width == size; height > max_size;
# do_pad=False, longest_edge=118, shortest_edge=100, image=120x100 -> 118x98
image_processor = ConditionalDetrImageProcessor(
size={"longest_edge": 118, "shortest_edge": 100},
do_pad=False,
)
inputs = image_processor(images=[image_3], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 98, 118]))
image_4 = torch.ones([128, 50, 3], dtype=torch.uint8)
# max size is set; height == size; width < max_size;
# do_pad=False, longest_edge=256, shortest_edge=50, image=50x128 -> 50x128
image_processor = ConditionalDetrImageProcessor(
size={"longest_edge": 256, "shortest_edge": 50},
do_pad=False,
)
inputs = image_processor(images=[image_4], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 128, 50]))
image_5 = torch.ones([50, 50, 3], dtype=torch.uint8)
# max size is set; height == width; width < max_size;
# do_pad=False, longest_edge=117, shortest_edge=50, image=50x50 -> 50x50
image_processor = ConditionalDetrImageProcessor(
size={"longest_edge": 117, "shortest_edge": 50},
do_pad=False,
)
inputs = image_processor(images=[image_5], return_tensors="pt")
self.assertEqual(inputs["pixel_values"].shape, torch.Size([1, 3, 50, 50]))

630
transformers/tests/models/conditional_detr/test_modeling_conditional_detr.py Normal file
View File

@@ -0,0 +1,630 @@
# 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 Conditional DETR model."""
import inspect
import math
import unittest
from functools import cached_property
from transformers import ConditionalDetrConfig, ResNetConfig, is_torch_available, is_vision_available
from transformers.testing_utils import require_timm, require_torch, require_vision, slow, torch_device
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
ConditionalDetrForObjectDetection,
ConditionalDetrForSegmentation,
ConditionalDetrModel,
)
if is_vision_available():
from PIL import Image
from transformers import ConditionalDetrImageProcessor
class ConditionalDetrModelTester:
def __init__(
self,
parent,
batch_size=8,
is_training=True,
use_labels=True,
hidden_size=32,
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,
min_size=200,
max_size=200,
n_targets=8,
num_labels=91,
):
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.min_size = min_size
self.max_size = max_size
self.n_targets = n_targets
self.num_labels = num_labels
# we also set the expected seq length for both encoder and decoder
self.encoder_seq_length = math.ceil(self.min_size / 32) * math.ceil(self.max_size / 32)
self.decoder_seq_length = self.num_queries
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.min_size, self.max_size])
pixel_mask = torch.ones([self.batch_size, self.min_size, self.max_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.min_size, self.max_size, device=torch_device)
labels.append(target)
config = self.get_config()
return config, pixel_values, pixel_mask, labels
def get_config(self):
resnet_config = ResNetConfig(
num_channels=3,
embeddings_size=10,
hidden_sizes=[10, 20, 30, 40],
depths=[1, 1, 2, 1],
hidden_act="relu",
num_labels=3,
out_features=["stage2", "stage3", "stage4"],
out_indices=[2, 3, 4],
)
return ConditionalDetrConfig(
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,
use_timm_backbone=False,
backbone_config=resnet_config,
backbone=None,
use_pretrained_backbone=False,
)
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_conditional_detr_model(self, config, pixel_values, pixel_mask, labels):
model = ConditionalDetrModel(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.decoder_seq_length, self.hidden_size)
)
def create_and_check_conditional_detr_object_detection_head_model(self, config, pixel_values, pixel_mask, labels):
model = ConditionalDetrForObjectDetection(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_torch
class ConditionalDetrModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
ConditionalDetrModel,
ConditionalDetrForObjectDetection,
ConditionalDetrForSegmentation,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = (
{"image-feature-extraction": ConditionalDetrModel, "object-detection": ConditionalDetrForObjectDetection}
if is_torch_available()
else {}
)
is_encoder_decoder = True
test_torchscript = False
test_pruning = False
test_head_masking = False
test_missing_keys = False
zero_init_hidden_state = True
test_torch_exportable = True
# 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__ in ["ConditionalDetrForObjectDetection", "ConditionalDetrForSegmentation"]:
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.min_size,
self.model_tester.max_size,
device=torch_device,
dtype=torch.float,
)
labels.append(target)
inputs_dict["labels"] = labels
return inputs_dict
def setUp(self):
self.model_tester = ConditionalDetrModelTester(self)
self.config_tester = ConfigTester(self, config_class=ConditionalDetrConfig, has_text_modality=False)
def test_config(self):
self.config_tester.run_common_tests()
def test_conditional_detr_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_conditional_detr_model(*config_and_inputs)
def test_conditional_detr_object_detection_head_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_conditional_detr_object_detection_head_model(*config_and_inputs)
# TODO: check if this works again for PyTorch 2.x.y
@unittest.skip(reason="Got `CUDA error: misaligned address` with PyTorch 2.0.0.")
def test_multi_gpu_data_parallel_forward(self):
pass
@unittest.skip(reason="Conditional DETR does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Conditional DETR does not use inputs_embeds")
def test_inputs_embeds_matches_input_ids(self):
pass
@unittest.skip(reason="Conditional DETR does not have a get_input_embeddings method")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="Conditional DETR is not a generative model")
def test_generate_without_input_ids(self):
pass
@unittest.skip(reason="Conditional DETR does not use token embeddings")
def test_resize_tokens_embeddings(self):
pass
@slow
@unittest.skip(reason="TODO Niels: fix me!")
def test_model_outputs_equivalence(self):
pass
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
decoder_seq_length = self.model_tester.decoder_seq_length
encoder_seq_length = self.model_tester.encoder_seq_length
decoder_key_length = self.model_tester.decoder_seq_length
encoder_key_length = self.model_tester.encoder_seq_length
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._from_config(config, attn_implementation="eager")
config = model.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 if config.is_encoder_decoder else outputs.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 if config.is_encoder_decoder else 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, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
if self.is_encoder_decoder:
correct_outlen = 6
# 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__ == "ConditionalDetrForObjectDetection":
correct_outlen += 1
# Panoptic Segmentation model returns pred_logits, pred_boxes, pred_masks
if model_class.__name__ == "ConditionalDetrForSegmentation":
correct_outlen += 2
if "past_key_values" in outputs:
correct_outlen += 1 # past_key_values have been returned
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, decoder_seq_length, decoder_key_length],
)
# 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,
decoder_seq_length,
encoder_key_length,
],
)
# 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 if config.is_encoder_decoder else 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, encoder_seq_length, encoder_key_length],
)
def test_retain_grad_hidden_states_attentions(self):
# removed retain_grad and grad on decoder_hidden_states, as queries don't require grad
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)
output = outputs[0]
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_auxiliary_loss(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.auxiliary_loss = True
# only test for object detection and segmentation model
for model_class in self.all_model_classes[1:]:
model = model_class(config)
model.to(torch_device)
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
outputs = model(**inputs)
self.assertIsNotNone(outputs.auxiliary_outputs)
self.assertEqual(len(outputs.auxiliary_outputs), self.model_tester.num_hidden_layers - 1)
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)
def test_different_timm_backbone(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
# let's pick a random timm backbone
config.backbone = "tf_mobilenetv3_small_075"
config.backbone_config = None
config.use_timm_backbone = True
config.backbone_kwargs = {"out_indices": [2, 3, 4]}
for model_class in self.all_model_classes:
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 model_class.__name__ == "ConditionalDetrForObjectDetection":
expected_shape = (
self.model_tester.batch_size,
self.model_tester.num_queries,
self.model_tester.num_labels,
)
self.assertEqual(outputs.logits.shape, expected_shape)
# Confirm out_indices was propagated to backbone
self.assertEqual(len(model.model.backbone.conv_encoder.intermediate_channel_sizes), 3)
elif model_class.__name__ == "ConditionalDetrForSegmentation":
# Confirm out_indices was propagated to backbone
self.assertEqual(len(model.conditional_detr.model.backbone.conv_encoder.intermediate_channel_sizes), 3)
else:
# Confirm out_indices was propagated to backbone
self.assertEqual(len(model.backbone.conv_encoder.intermediate_channel_sizes), 3)
self.assertTrue(outputs)
@require_timm
def test_hf_backbone(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
# Load a pretrained HF checkpoint as backbone
config.backbone = "microsoft/resnet-18"
config.backbone_config = None
config.use_timm_backbone = False
config.use_pretrained_backbone = True
config.backbone_kwargs = {"out_indices": [2, 3, 4]}
for model_class in self.all_model_classes:
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 model_class.__name__ == "ConditionalDetrForObjectDetection":
expected_shape = (
self.model_tester.batch_size,
self.model_tester.num_queries,
self.model_tester.num_labels,
)
self.assertEqual(outputs.logits.shape, expected_shape)
# Confirm out_indices was propagated to backbone
self.assertEqual(len(model.model.backbone.conv_encoder.intermediate_channel_sizes), 3)
elif model_class.__name__ == "ConditionalDetrForSegmentation":
# Confirm out_indices was propagated to backbone
self.assertEqual(len(model.conditional_detr.model.backbone.conv_encoder.intermediate_channel_sizes), 3)
else:
# Confirm out_indices was propagated to backbone
self.assertEqual(len(model.backbone.conv_encoder.intermediate_channel_sizes), 3)
self.assertTrue(outputs)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
configs_no_init.init_xavier_std = 1e9
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 "bbox_attention" in name and "bias" not in name:
self.assertLess(
100000,
abs(param.data.max().item()),
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
else:
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_timm
@require_vision
@slow
class ConditionalDetrModelIntegrationTests(unittest.TestCase):
@cached_property
def default_image_processor(self):
return (
ConditionalDetrImageProcessor.from_pretrained("microsoft/conditional-detr-resnet-50")
if is_vision_available()
else None
)
def test_inference_no_head(self):
model = ConditionalDetrModel.from_pretrained("microsoft/conditional-detr-resnet-50").to(torch_device)
image_processor = self.default_image_processor
image = prepare_img()
encoding = image_processor(images=image, return_tensors="pt").to(torch_device)
with torch.no_grad():
outputs = model(**encoding)
expected_shape = torch.Size((1, 300, 256))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor(
[
[0.4223, 0.7474, 0.8760],
[0.6397, -0.2727, 0.7126],
[-0.3089, 0.7643, 0.9529],
]
).to(torch_device)
torch.testing.assert_close(outputs.last_hidden_state[0, :3, :3], expected_slice, rtol=2e-4, atol=2e-4)
def test_inference_object_detection_head(self):
model = ConditionalDetrForObjectDetection.from_pretrained("microsoft/conditional-detr-resnet-50").to(
torch_device
)
image_processor = self.default_image_processor
image = prepare_img()
encoding = image_processor(images=image, return_tensors="pt").to(torch_device)
pixel_values = encoding["pixel_values"].to(torch_device)
pixel_mask = encoding["pixel_mask"].to(torch_device)
with torch.no_grad():
outputs = model(pixel_values, pixel_mask)
# verify logits + box predictions
expected_shape_logits = torch.Size((1, model.config.num_queries, model.config.num_labels))
self.assertEqual(outputs.logits.shape, expected_shape_logits)
expected_slice_logits = torch.tensor(
[
[-10.4371, -5.7565, -8.6765],
[-10.5413, -5.8700, -8.0589],
[-10.6824, -6.3477, -8.3927],
]
).to(torch_device)
torch.testing.assert_close(outputs.logits[0, :3, :3], expected_slice_logits, rtol=2e-4, atol=2e-4)
expected_shape_boxes = torch.Size((1, model.config.num_queries, 4))
self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
expected_slice_boxes = torch.tensor(
[
[0.7733, 0.6576, 0.4496],
[0.5171, 0.1184, 0.9095],
[0.8846, 0.5647, 0.2486],
]
).to(torch_device)
torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_slice_boxes, rtol=2e-4, atol=2e-4)
# verify postprocessing
results = image_processor.post_process_object_detection(
outputs, threshold=0.3, target_sizes=[image.size[::-1]]
)[0]
expected_scores = torch.tensor([0.8330, 0.8315, 0.8039, 0.6829, 0.5354]).to(torch_device)
expected_labels = [75, 17, 17, 75, 63]
expected_slice_boxes = torch.tensor([38.3089, 72.1023, 177.6292, 118.4514]).to(torch_device)
self.assertEqual(len(results["scores"]), 5)
torch.testing.assert_close(results["scores"], expected_scores, rtol=2e-4, atol=2e-4)
self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
torch.testing.assert_close(results["boxes"][0, :], expected_slice_boxes)