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/yolos/__init__.py Normal file
View File

612
transformers/tests/models/yolos/test_image_processing_yolos.py Normal file
View File

@@ -0,0 +1,612 @@
# Copyright 2021 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 parameterized import parameterized
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 YolosImageProcessor
if is_torchvision_available():
from transformers import YolosImageProcessorFast
class YolosImageProcessingTester:
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 YolosImageProcessor,
assuming do_resize is set to True with a scalar size.
"""
if not batched:
image = image_inputs[0]
if isinstance(image, Image.Image):
width, height = image.size
elif isinstance(image, np.ndarray):
height, width = image.shape[0], image.shape[1]
else:
height, width = image.shape[1], image.shape[2]
size = self.size["shortest_edge"]
max_size = self.size.get("longest_edge", None)
if max_size is not None:
min_original_size = float(min((height, width)))
max_original_size = float(max((height, width)))
if max_original_size / min_original_size * size > max_size:
size = int(round(max_size * min_original_size / max_original_size))
if width <= height and width != size:
height = int(size * height / width)
width = size
elif height < width and height != size:
width = int(size * width / height)
height = size
width_mod = width % 16
height_mod = height % 16
expected_width = width - width_mod
expected_height = height - height_mod
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 YolosImageProcessingTest(AnnotationFormatTestMixin, ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = YolosImageProcessor if is_vision_available() else None
fast_image_processing_class = YolosImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = YolosImageProcessingTester(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)
def test_equivalence_padding(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 = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
# Test whether the method "pad" and calling the image processor return the same tensors
encoded_images_with_method = image_processing_1.pad(image_inputs, return_tensors="pt")
encoded_images = image_processing_2(image_inputs, return_tensors="pt")
torch.testing.assert_close(
encoded_images_with_method["pixel_values"], encoded_images["pixel_values"], rtol=1e-4, atol=1e-4
)
@parameterized.expand(
[
((3, 100, 1500), 1333, 800),
((3, 400, 400), 1333, 800),
((3, 1500, 1500), 1333, 800),
((3, 800, 1333), 1333, 800),
((3, 1333, 800), 1333, 800),
((3, 800, 800), 400, 400),
]
)
def test_resize_max_size_respected(self, image_size, longest_edge, shortest_edge):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
# create torch tensors as image
image = torch.randint(0, 256, image_size, dtype=torch.uint8)
processed_image = image_processor(
image,
size={"longest_edge": longest_edge, "shortest_edge": shortest_edge},
do_pad=False,
return_tensors="pt",
)["pixel_values"]
shape = list(processed_image.shape[-2:])
max_size, min_size = max(shape), min(shape)
self.assertTrue(max_size <= 1333, f"Expected max_size <= 1333, got image shape {shape}")
self.assertTrue(min_size <= 800, f"Expected min_size <= 800, got image shape {shape}")
@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("hustvl/yolos-small")
encoding = image_processing(images=image, annotations=target, return_tensors="pt")
# verify pixel values
expected_shape = torch.Size([1, 3, 800, 1056])
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([5832.7256, 11144.6689, 484763.2500, 829269.8125, 146579.4531, 164177.6250])
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, 1056])
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, 1056])
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([146591.5000, 163974.2500, 480092.2500, 11187.0000, 5824.5000, 7562.5000])
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 = 815161
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, 1056])
torch.testing.assert_close(encoding["labels"][0]["size"], expected_size)
# Output size is slight different from DETR as yolos takes mod of 16
@slow
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, 1056
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(
[
[0.4169, 0.2765, 0.0458, 0.2215],
[0.1284, 0.2016, 0.1576, 0.0940],
[0.3792, 0.4933, 0.7559, 0.9865],
[0.3794, 0.5002, 0.7563, 0.9955],
[0.1990, 0.5456, 0.3566, 0.8646],
[0.5845, 0.4115, 0.3462, 0.7161],
]
)
torch.testing.assert_close(encoding["labels"][0]["boxes"], expected_boxes_0, rtol=1e-3, atol=1e-3)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, rtol=1e-3, atol=1e-3)
# Check the masks have also been padded
self.assertEqual(encoding["labels"][0]["masks"].shape, torch.Size([6, 800, 1056]))
self.assertEqual(encoding["labels"][1]["masks"].shape, torch.Size([6, 800, 1056]))
# 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, rtol=1, atol=1)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, rtol=1, atol=1)
# Output size is slight different from DETR as yolos takes mod of 16
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]
# encode them
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class()
image_processing = YolosImageProcessor(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, 1056
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(
[
[0.1591, 0.3262, 0.2841, 0.5175],
[0.4678, 0.2463, 0.2746, 0.4275],
[0.3030, 0.2956, 0.6042, 0.5913],
[0.1023, 0.1200, 0.1250, 0.0550],
[0.3329, 0.1656, 0.0350, 0.1312],
[0.3026, 0.2994, 0.6051, 0.5987],
]
)
torch.testing.assert_close(encoding["labels"][0]["boxes"], expected_boxes_0, rtol=1e-3, atol=1e-3)
torch.testing.assert_close(encoding["labels"][1]["boxes"], expected_boxes_1, rtol=1e-3, atol=1e-3)
# Check the masks have also been padded
self.assertEqual(encoding["labels"][0]["masks"].shape, torch.Size([6, 800, 1056]))
self.assertEqual(encoding["labels"][1]["masks"].shape, torch.Size([6, 800, 1056]))
# 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->Yolos
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]))

380
transformers/tests/models/yolos/test_modeling_yolos.py Normal file
View File

@@ -0,0 +1,380 @@
# 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 YOLOS model."""
import unittest
from functools import cached_property
from transformers import YolosConfig
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from transformers.utils import is_torch_available, is_vision_available
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, floats_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import YolosForObjectDetection, YolosModel
if is_vision_available():
from PIL import Image
from transformers import AutoImageProcessor
class YolosModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=[30, 30],
patch_size=2,
num_channels=3,
is_training=True,
use_labels=True,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
type_sequence_label_size=10,
initializer_range=0.02,
num_labels=3,
scope=None,
n_targets=8,
num_detection_tokens=10,
attn_implementation="eager",
):
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.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.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.scope = scope
self.n_targets = n_targets
self.num_detection_tokens = num_detection_tokens
self.attn_implementation = attn_implementation
# we set the expected sequence length (which is used in several tests)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token) + num_detection_tokens
num_patches = (image_size[1] // patch_size) * (image_size[0] // patch_size)
self.expected_seq_len = num_patches + 1 + self.num_detection_tokens
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size[0], self.image_size[1]])
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)
labels.append(target)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return YolosConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
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,
num_detection_tokens=self.num_detection_tokens,
num_labels=self.num_labels,
attn_implementation=self.attn_implementation,
)
def create_and_check_model(self, config, pixel_values, labels):
model = YolosModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.last_hidden_state.shape, (self.batch_size, self.expected_seq_len, self.hidden_size)
)
def create_and_check_for_object_detection(self, config, pixel_values, labels):
model = YolosForObjectDetection(config)
model.to(torch_device)
model.eval()
result = model(pixel_values=pixel_values)
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_detection_tokens, self.num_labels + 1))
self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_detection_tokens, 4))
result = model(pixel_values=pixel_values, labels=labels)
self.parent.assertEqual(result.loss.shape, ())
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_detection_tokens, self.num_labels + 1))
self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_detection_tokens, 4))
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 YolosModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as YOLOS does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (YolosModel, YolosForObjectDetection) if is_torch_available() else ()
pipeline_model_mapping = (
{"image-feature-extraction": YolosModel, "object-detection": YolosForObjectDetection}
if is_torch_available()
else {}
)
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
test_torchscript = False
test_torch_exportable = True
# special case for head model
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__ == "YolosForObjectDetection":
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
)
labels.append(target)
inputs_dict["labels"] = labels
return inputs_dict
def setUp(self):
self.model_tester = YolosModelTester(self)
self.config_tester = ConfigTester(self, config_class=YolosConfig, has_text_modality=False, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="YOLOS does not use inputs_embeds")
def test_inputs_embeds(self):
pass
def test_model_get_set_embeddings(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, nn.Linear))
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
# in YOLOS, the seq_len is different
seq_len = self.model_tester.expected_seq_len
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.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.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))
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, 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)
# YOLOS has a different seq_length
seq_length = self.model_tester.expected_seq_len
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, 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_for_object_detection(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_object_detection(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
model_name = "hustvl/yolos-small"
model = YolosModel.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
class YolosModelIntegrationTest(unittest.TestCase):
@cached_property
def default_image_processor(self):
return AutoImageProcessor.from_pretrained("hustvl/yolos-small") if is_vision_available() else None
@slow
def test_inference_object_detection_head(self):
model = YolosForObjectDetection.from_pretrained("hustvl/yolos-small").to(torch_device)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(inputs.pixel_values)
# verify outputs
expected_shape = torch.Size((1, 100, 92))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice_logits = torch.tensor(
[[-23.7219, -10.3165, -14.9083], [-41.5429, -15.2403, -24.1478], [-29.3909, -12.7173, -19.4650]],
device=torch_device,
)
expected_slice_boxes = torch.tensor(
[[0.2536, 0.5449, 0.4643], [0.2037, 0.7735, 0.3672], [0.7692, 0.4056, 0.4549]], device=torch_device
)
torch.testing.assert_close(outputs.logits[0, :3, :3], expected_slice_logits, rtol=1e-4, atol=1e-4)
torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_slice_boxes, rtol=1e-4, 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([0.9991, 0.9801, 0.9978, 0.9875, 0.9848]).to(torch_device)
expected_labels = [75, 75, 17, 63, 17]
expected_slice_boxes = torch.tensor([331.8438, 80.5440, 369.9546, 188.0579]).to(torch_device)
self.assertEqual(len(results["scores"]), 5)
torch.testing.assert_close(results["scores"], expected_scores, rtol=1e-4, atol=1e-4)
self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
torch.testing.assert_close(results["boxes"][0, :], expected_slice_boxes)