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/maskformer/__init__.py Normal file
View File

592
transformers/tests/models/maskformer/test_image_processing_maskformer.py Normal file
View File

@@ -0,0 +1,592 @@
# Copyright 2022 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
from datasets import load_dataset
from huggingface_hub import hf_hub_download
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
if is_torch_available():
import torch
if is_vision_available():
from transformers import MaskFormerImageProcessor
from transformers.models.maskformer.image_processing_maskformer import binary_mask_to_rle
from transformers.models.maskformer.modeling_maskformer import MaskFormerForInstanceSegmentationOutput
if is_torchvision_available():
from transformers import MaskFormerImageProcessorFast
if is_vision_available():
from PIL import Image
class MaskFormerImageProcessingTester:
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
min_resolution=30,
max_resolution=400,
size=None,
do_resize=True,
do_normalize=True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
num_labels=10,
do_reduce_labels=True,
ignore_index=255,
):
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 = {"shortest_edge": 32, "longest_edge": 1333} if size is None else size
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
self.size_divisor = 0
# for the post_process_functions
self.batch_size = 2
self.num_queries = 3
self.num_classes = 2
self.height = 3
self.width = 4
self.num_labels = num_labels
self.do_reduce_labels = do_reduce_labels
self.ignore_index = ignore_index
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,
"size_divisor": self.size_divisor,
"num_labels": self.num_labels,
"do_reduce_labels": self.do_reduce_labels,
"ignore_index": self.ignore_index,
}
def get_expected_values(self, image_inputs, batched=False):
"""
This function computes the expected height and width when providing images to MaskFormerImageProcessor,
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 get_fake_maskformer_outputs(self):
return MaskFormerForInstanceSegmentationOutput(
# +1 for null class
class_queries_logits=torch.randn((self.batch_size, self.num_queries, self.num_classes + 1)),
masks_queries_logits=torch.randn((self.batch_size, self.num_queries, self.height, self.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,
)
# Copied from transformers.tests.models.beit.test_image_processing_beit.prepare_semantic_single_inputs
def prepare_semantic_single_inputs():
ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
example = ds[0]
return example["image"], example["map"]
# Copied from transformers.tests.models.beit.test_image_processing_beit.prepare_semantic_batch_inputs
def prepare_semantic_batch_inputs():
ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
return list(ds["image"][:2]), list(ds["map"][:2])
@require_torch
@require_vision
class MaskFormerImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = MaskFormerImageProcessor if (is_vision_available() and is_torch_available()) else None
fast_image_processing_class = (
MaskFormerImageProcessorFast if (is_vision_available() and is_torchvision_available()) else None
)
def setUp(self):
super().setUp()
self.image_processor_tester = MaskFormerImageProcessingTester(self)
@property
def image_processor_dict(self):
return self.image_processor_tester.prepare_image_processor_dict()
def test_image_processor_properties(self):
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
self.assertTrue(hasattr(image_processing, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "ignore_index"))
self.assertTrue(hasattr(image_processing, "num_labels"))
def comm_get_image_processing_inputs(
self, with_segmentation_maps=False, is_instance_map=False, segmentation_type="np"
):
image_processing = self.image_processing_class(**self.image_processor_dict)
# prepare image and target
num_labels = self.image_processor_tester.num_labels
annotations = None
instance_id_to_semantic_id = None
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
if with_segmentation_maps:
high = num_labels
if is_instance_map:
labels_expanded = list(range(num_labels)) * 2
instance_id_to_semantic_id = dict(enumerate(labels_expanded))
annotations = [
np.random.randint(0, high * 2, (img.size[1], img.size[0])).astype(np.uint8) for img in image_inputs
]
if segmentation_type == "pil":
annotations = [Image.fromarray(annotation) for annotation in annotations]
inputs = image_processing(
image_inputs,
annotations,
return_tensors="pt",
instance_id_to_semantic_id=instance_id_to_semantic_id,
pad_and_return_pixel_mask=True,
)
return inputs
def test_with_size_divisor(self):
size_divisors = [8, 16, 32]
weird_input_sizes = [(407, 802), (582, 1094)]
for image_processing_class in self.image_processor_list:
for size_divisor in size_divisors:
image_processor_dict = {**self.image_processor_dict, **{"size_divisor": size_divisor}}
image_processing = image_processing_class(**image_processor_dict)
for weird_input_size in weird_input_sizes:
inputs = image_processing([np.ones((3, *weird_input_size))], return_tensors="pt")
pixel_values = inputs["pixel_values"]
# check if divisible
self.assertTrue((pixel_values.shape[-1] % size_divisor) == 0)
self.assertTrue((pixel_values.shape[-2] % size_divisor) == 0)
def test_call_with_segmentation_maps(self):
def common(is_instance_map=False, segmentation_type=None):
inputs = self.comm_get_image_processing_inputs(
with_segmentation_maps=True, is_instance_map=is_instance_map, segmentation_type=segmentation_type
)
mask_labels = inputs["mask_labels"]
class_labels = inputs["class_labels"]
pixel_values = inputs["pixel_values"]
# check the batch_size
for mask_label, class_label in zip(mask_labels, class_labels):
self.assertEqual(mask_label.shape[0], class_label.shape[0])
# this ensure padding has happened
self.assertEqual(mask_label.shape[1:], pixel_values.shape[2:])
common()
common(is_instance_map=True)
common(is_instance_map=False, segmentation_type="pil")
common(is_instance_map=True, segmentation_type="pil")
def test_integration_instance_segmentation(self):
# load 2 images and corresponding annotations from the hub
repo_id = "nielsr/image-segmentation-toy-data"
image1 = Image.open(
hf_hub_download(repo_id=repo_id, filename="instance_segmentation_image_1.png", repo_type="dataset")
)
image2 = Image.open(
hf_hub_download(repo_id=repo_id, filename="instance_segmentation_image_2.png", repo_type="dataset")
)
annotation1 = Image.open(
hf_hub_download(repo_id=repo_id, filename="instance_segmentation_annotation_1.png", repo_type="dataset")
)
annotation2 = Image.open(
hf_hub_download(repo_id=repo_id, filename="instance_segmentation_annotation_2.png", repo_type="dataset")
)
# get instance segmentations and instance-to-segmentation mappings
def get_instance_segmentation_and_mapping(annotation):
instance_seg = np.array(annotation)[:, :, 1]
class_id_map = np.array(annotation)[:, :, 0]
class_labels = np.unique(class_id_map)
# create mapping between instance IDs and semantic category IDs
inst2class = {}
for label in class_labels:
instance_ids = np.unique(instance_seg[class_id_map == label])
inst2class.update(dict.fromkeys(instance_ids, label))
return instance_seg, inst2class
instance_seg1, inst2class1 = get_instance_segmentation_and_mapping(annotation1)
instance_seg2, inst2class2 = get_instance_segmentation_and_mapping(annotation2)
# create a image processor
image_processing = MaskFormerImageProcessor(do_reduce_labels=True, ignore_index=255, size=(512, 512))
# prepare the images and annotations
inputs = image_processing(
[image1, image2],
[instance_seg1, instance_seg2],
instance_id_to_semantic_id=[inst2class1, inst2class2],
return_tensors="pt",
)
# verify the pixel values and pixel mask
self.assertEqual(inputs["pixel_values"].shape, (2, 3, 512, 512))
self.assertEqual(inputs["pixel_mask"].shape, (2, 512, 512))
# verify the class labels
self.assertEqual(len(inputs["class_labels"]), 2)
torch.testing.assert_close(inputs["class_labels"][0], torch.tensor([30, 55]))
torch.testing.assert_close(inputs["class_labels"][1], torch.tensor([4, 4, 23, 55]))
# verify the mask labels
self.assertEqual(len(inputs["mask_labels"]), 2)
self.assertEqual(inputs["mask_labels"][0].shape, (2, 512, 512))
self.assertEqual(inputs["mask_labels"][1].shape, (4, 512, 512))
self.assertEqual(inputs["mask_labels"][0].sum().item(), 41527.0)
self.assertEqual(inputs["mask_labels"][1].sum().item(), 26259.0)
def test_integration_semantic_segmentation(self):
# load 2 images and corresponding semantic annotations from the hub
repo_id = "nielsr/image-segmentation-toy-data"
image1 = Image.open(
hf_hub_download(repo_id=repo_id, filename="semantic_segmentation_image_1.png", repo_type="dataset")
)
image2 = Image.open(
hf_hub_download(repo_id=repo_id, filename="semantic_segmentation_image_2.png", repo_type="dataset")
)
annotation1 = Image.open(
hf_hub_download(repo_id=repo_id, filename="semantic_segmentation_annotation_1.png", repo_type="dataset")
)
annotation2 = Image.open(
hf_hub_download(repo_id=repo_id, filename="semantic_segmentation_annotation_2.png", repo_type="dataset")
)
# create a image processor
image_processing = MaskFormerImageProcessor(do_reduce_labels=True, ignore_index=255, size=(512, 512))
# prepare the images and annotations
inputs = image_processing(
[image1, image2],
[annotation1, annotation2],
return_tensors="pt",
)
# verify the pixel values and pixel mask
self.assertEqual(inputs["pixel_values"].shape, (2, 3, 512, 512))
self.assertEqual(inputs["pixel_mask"].shape, (2, 512, 512))
# verify the class labels
self.assertEqual(len(inputs["class_labels"]), 2)
torch.testing.assert_close(inputs["class_labels"][0], torch.tensor([2, 4, 60]))
torch.testing.assert_close(inputs["class_labels"][1], torch.tensor([0, 3, 7, 8, 15, 28, 30, 143]))
# verify the mask labels
self.assertEqual(len(inputs["mask_labels"]), 2)
self.assertEqual(inputs["mask_labels"][0].shape, (3, 512, 512))
self.assertEqual(inputs["mask_labels"][1].shape, (8, 512, 512))
self.assertEqual(inputs["mask_labels"][0].sum().item(), 170200.0)
self.assertEqual(inputs["mask_labels"][1].sum().item(), 257036.0)
def test_integration_panoptic_segmentation(self):
# load 2 images and corresponding panoptic annotations from the hub
dataset = load_dataset("nielsr/ade20k-panoptic-demo")
image1 = dataset["train"][0]["image"]
image2 = dataset["train"][1]["image"]
segments_info1 = dataset["train"][0]["segments_info"]
segments_info2 = dataset["train"][1]["segments_info"]
annotation1 = dataset["train"][0]["label"]
annotation2 = dataset["train"][1]["label"]
def rgb_to_id(color):
if isinstance(color, np.ndarray) and len(color.shape) == 3:
if color.dtype == np.uint8:
color = color.astype(np.int32)
return color[:, :, 0] + 256 * color[:, :, 1] + 256 * 256 * color[:, :, 2]
return int(color[0] + 256 * color[1] + 256 * 256 * color[2])
def create_panoptic_map(annotation, segments_info):
annotation = np.array(annotation)
# convert RGB to segment IDs per pixel
# 0 is the "ignore" label, for which we don't need to make binary masks
panoptic_map = rgb_to_id(annotation)
# create mapping between segment IDs and semantic classes
inst2class = {segment["id"]: segment["category_id"] for segment in segments_info}
return panoptic_map, inst2class
panoptic_map1, inst2class1 = create_panoptic_map(annotation1, segments_info1)
panoptic_map2, inst2class2 = create_panoptic_map(annotation2, segments_info2)
# create a image processor
image_processing = MaskFormerImageProcessor(ignore_index=0, do_resize=False)
# prepare the images and annotations
pixel_values_list = [np.moveaxis(np.array(image1), -1, 0), np.moveaxis(np.array(image2), -1, 0)]
inputs = image_processing.encode_inputs(
pixel_values_list,
[panoptic_map1, panoptic_map2],
instance_id_to_semantic_id=[inst2class1, inst2class2],
return_tensors="pt",
)
# verify the pixel values and pixel mask
self.assertEqual(inputs["pixel_values"].shape, (2, 3, 512, 711))
self.assertEqual(inputs["pixel_mask"].shape, (2, 512, 711))
# verify the class labels
self.assertEqual(len(inputs["class_labels"]), 2)
expected_class_labels = torch.tensor([4, 17, 32, 42, 42, 42, 42, 42, 42, 42, 32, 12, 12, 12, 12, 12, 42, 42, 12, 12, 12, 42, 12, 12, 12, 12, 12, 3, 12, 12, 12, 12, 42, 42, 42, 12, 42, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 5, 12, 12, 12, 12, 12, 12, 12, 0, 43, 43, 43, 96, 43, 104, 43, 31, 125, 31, 125, 138, 87, 125, 149, 138, 125, 87, 87]) # fmt: skip
torch.testing.assert_close(inputs["class_labels"][0], torch.tensor(expected_class_labels))
expected_class_labels = torch.tensor([19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 67, 82, 19, 19, 17, 19, 19, 19, 19, 19, 19, 19, 19, 19, 12, 12, 42, 12, 12, 12, 12, 3, 14, 12, 12, 12, 12, 12, 12, 12, 12, 14, 5, 12, 12, 0, 115, 43, 43, 115, 43, 43, 43, 8, 8, 8, 138, 138, 125, 143]) # fmt: skip
torch.testing.assert_close(inputs["class_labels"][1], expected_class_labels)
# verify the mask labels
self.assertEqual(len(inputs["mask_labels"]), 2)
self.assertEqual(inputs["mask_labels"][0].shape, (79, 512, 711))
self.assertEqual(inputs["mask_labels"][1].shape, (61, 512, 711))
self.assertEqual(inputs["mask_labels"][0].sum().item(), 315193.0)
self.assertEqual(inputs["mask_labels"][1].sum().item(), 350747.0)
def test_binary_mask_to_rle(self):
fake_binary_mask = np.zeros((20, 50))
fake_binary_mask[0, 20:] = 1
fake_binary_mask[1, :15] = 1
fake_binary_mask[5, :10] = 1
rle = binary_mask_to_rle(fake_binary_mask)
self.assertEqual(len(rle), 4)
self.assertEqual(rle[0], 21)
self.assertEqual(rle[1], 45)
def test_post_process_segmentation(self):
for image_processing_class in self.image_processor_list:
feature_extractor = image_processing_class(num_labels=self.image_processor_tester.num_classes)
outputs = self.image_processor_tester.get_fake_maskformer_outputs()
segmentation = feature_extractor.post_process_segmentation(outputs)
self.assertEqual(
segmentation.shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.num_classes,
self.image_processor_tester.height,
self.image_processor_tester.width,
),
)
target_size = (1, 4)
segmentation = feature_extractor.post_process_segmentation(outputs, target_size=target_size)
self.assertEqual(
segmentation.shape,
(self.image_processor_tester.batch_size, self.image_processor_tester.num_classes, *target_size),
)
def test_post_process_semantic_segmentation(self):
for image_processing_class in self.image_processor_list:
feature_extractor = image_processing_class(num_labels=self.image_processor_tester.num_classes)
outputs = self.image_processor_tester.get_fake_maskformer_outputs()
segmentation = feature_extractor.post_process_semantic_segmentation(outputs)
self.assertEqual(len(segmentation), self.image_processor_tester.batch_size)
self.assertEqual(
segmentation[0].shape,
(
self.image_processor_tester.height,
self.image_processor_tester.width,
),
)
target_sizes = [(1, 4) for i in range(self.image_processor_tester.batch_size)]
segmentation = feature_extractor.post_process_semantic_segmentation(outputs, target_sizes=target_sizes)
self.assertEqual(segmentation[0].shape, target_sizes[0])
def test_post_process_instance_segmentation(self):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(num_labels=self.image_processor_tester.num_classes)
outputs = self.image_processor_tester.get_fake_maskformer_outputs()
segmentation = image_processor.post_process_instance_segmentation(outputs, threshold=0)
self.assertTrue(len(segmentation) == self.image_processor_tester.batch_size)
for el in segmentation:
self.assertTrue("segmentation" in el)
self.assertTrue("segments_info" in el)
self.assertEqual(type(el["segments_info"]), list)
self.assertEqual(
el["segmentation"].shape, (self.image_processor_tester.height, self.image_processor_tester.width)
)
segmentation = image_processor.post_process_instance_segmentation(
outputs, threshold=0, return_binary_maps=True
)
self.assertTrue(len(segmentation) == self.image_processor_tester.batch_size)
for el in segmentation:
self.assertTrue("segmentation" in el)
self.assertTrue("segments_info" in el)
self.assertEqual(type(el["segments_info"]), list)
self.assertEqual(len(el["segmentation"].shape), 3)
self.assertEqual(
el["segmentation"].shape[1:],
(self.image_processor_tester.height, self.image_processor_tester.width),
)
def test_post_process_panoptic_segmentation(self):
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(num_labels=self.image_processor_tester.num_classes)
outputs = self.image_processor_tester.get_fake_maskformer_outputs()
segmentation = image_processing.post_process_panoptic_segmentation(outputs, threshold=0)
self.assertTrue(len(segmentation) == self.image_processor_tester.batch_size)
for el in segmentation:
self.assertTrue("segmentation" in el)
self.assertTrue("segments_info" in el)
self.assertEqual(type(el["segments_info"]), list)
self.assertEqual(
el["segmentation"].shape, (self.image_processor_tester.height, self.image_processor_tester.width)
)
def test_post_process_label_fusing(self):
for image_processing_class in self.image_processor_list:
image_processor = self.image_processing_class(num_labels=self.image_processor_tester.num_classes)
outputs = self.image_processor_tester.get_fake_maskformer_outputs()
segmentation = image_processor.post_process_panoptic_segmentation(
outputs, threshold=0, mask_threshold=0, overlap_mask_area_threshold=0
)
unfused_segments = [el["segments_info"] for el in segmentation]
fused_segmentation = image_processor.post_process_panoptic_segmentation(
outputs, threshold=0, mask_threshold=0, overlap_mask_area_threshold=0, label_ids_to_fuse={1}
)
fused_segments = [el["segments_info"] for el in fused_segmentation]
for el_unfused, el_fused in zip(unfused_segments, fused_segments):
if len(el_unfused) == 0:
self.assertEqual(len(el_unfused), len(el_fused))
continue
# Get number of segments to be fused
fuse_targets = [1 for el in el_unfused if el["label_id"] in {1}]
num_to_fuse = 0 if len(fuse_targets) == 0 else sum(fuse_targets) - 1
# Expected number of segments after fusing
expected_num_segments = max([el["id"] for el in el_unfused]) - num_to_fuse
num_segments_fused = max([el["id"] for el in el_fused])
self.assertEqual(num_segments_fused, expected_num_segments)
def test_slow_fast_equivalence(self):
if not self.test_slow_image_processor or not self.test_fast_image_processor:
self.skipTest(reason="Skipping slow/fast equivalence test")
if self.image_processing_class is None or self.fast_image_processing_class is None:
self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
dummy_image, dummy_map = prepare_semantic_single_inputs()
image_processor_slow = self.image_processing_class(**self.image_processor_dict)
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
image_encoding_slow = image_processor_slow(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
image_encoding_fast = image_processor_fast(dummy_image, segmentation_maps=dummy_map, return_tensors="pt")
self._assert_slow_fast_tensors_equivalence(image_encoding_slow.pixel_values, image_encoding_fast.pixel_values)
for mask_label_slow, mask_label_fast in zip(image_encoding_slow.mask_labels, image_encoding_fast.mask_labels):
self._assert_slow_fast_tensors_equivalence(mask_label_slow, mask_label_fast)
for class_label_slow, class_label_fast in zip(
image_encoding_slow.class_labels, image_encoding_fast.class_labels
):
self._assert_slow_fast_tensors_equivalence(class_label_slow.float(), class_label_fast.float())
def test_slow_fast_equivalence_batched(self):
if not self.test_slow_image_processor or not self.test_fast_image_processor:
self.skipTest(reason="Skipping slow/fast equivalence test")
if self.image_processing_class is None or self.fast_image_processing_class is None:
self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
self.skipTest(
reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
)
dummy_images, dummy_maps = prepare_semantic_batch_inputs()
image_processor_slow = self.image_processing_class(**self.image_processor_dict)
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
encoding_slow = image_processor_slow(dummy_images, segmentation_maps=dummy_maps, return_tensors="pt")
encoding_fast = image_processor_fast(dummy_images, segmentation_maps=dummy_maps, return_tensors="pt")
self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values, encoding_fast.pixel_values)
for mask_label_slow, mask_label_fast in zip(encoding_slow.mask_labels, encoding_fast.mask_labels):
self._assert_slow_fast_tensors_equivalence(mask_label_slow, mask_label_fast)
for class_label_slow, class_label_fast in zip(encoding_slow.class_labels, encoding_fast.class_labels):
self._assert_slow_fast_tensors_equivalence(class_label_slow.float(), class_label_fast.float())

709
transformers/tests/models/maskformer/test_modeling_maskformer.py Normal file
View File

@@ -0,0 +1,709 @@
# 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 MaskFormer model."""
import copy
import unittest
from functools import cached_property
import numpy as np
from tests.test_modeling_common import floats_tensor
from transformers import DetrConfig, MaskFormerConfig, SwinConfig, is_torch_available, is_vision_available
from transformers.testing_utils import (
Expectations,
require_timm,
require_torch,
require_torch_accelerator,
require_torch_fp16,
require_torch_multi_gpu,
require_vision,
slow,
torch_device,
)
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
import torch.nn.functional as F
from transformers import MaskFormerForInstanceSegmentation, MaskFormerModel
if is_vision_available():
from transformers import MaskFormerImageProcessor
if is_vision_available():
from PIL import Image
class MaskFormerModelTester:
def __init__(
self,
parent,
batch_size=2,
is_training=True,
use_auxiliary_loss=False,
num_queries=10,
num_channels=3,
min_size=32 * 4,
max_size=32 * 6,
num_labels=4,
mask_feature_size=32,
num_hidden_layers=2,
num_attention_heads=2,
):
self.parent = parent
self.batch_size = batch_size
self.is_training = is_training
self.use_auxiliary_loss = use_auxiliary_loss
self.num_queries = num_queries
self.num_channels = num_channels
self.min_size = min_size
self.max_size = max_size
self.num_labels = num_labels
self.mask_feature_size = mask_feature_size
# This is passed to the decoder config. We add it to the model tester here for testing
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.min_size, self.max_size]).to(
torch_device
)
pixel_mask = torch.ones([self.batch_size, self.min_size, self.max_size], device=torch_device)
mask_labels = (
torch.rand([self.batch_size, self.num_labels, self.min_size, self.max_size], device=torch_device) > 0.5
).float()
class_labels = (torch.rand((self.batch_size, self.num_labels), device=torch_device) > 0.5).long()
config = self.get_config()
return config, pixel_values, pixel_mask, mask_labels, class_labels
def get_config(self):
return MaskFormerConfig.from_backbone_and_decoder_configs(
backbone_config=SwinConfig(
depths=[1, 1, 1, 1],
embed_dim=16,
hidden_size=32,
num_heads=[1, 1, 2, 2],
),
backbone=None,
decoder_config=DetrConfig(
decoder_ffn_dim=64,
decoder_layers=self.num_hidden_layers,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=64,
encoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
num_queries=self.num_queries,
d_model=self.mask_feature_size,
),
mask_feature_size=self.mask_feature_size,
fpn_feature_size=self.mask_feature_size,
num_channels=self.num_channels,
num_labels=self.num_labels,
)
def prepare_config_and_inputs_for_common(self):
config, pixel_values, pixel_mask, _, _ = self.prepare_config_and_inputs()
inputs_dict = {"pixel_values": pixel_values, "pixel_mask": pixel_mask}
return config, inputs_dict
def check_output_hidden_state(self, output, config):
encoder_hidden_states = output.encoder_hidden_states
pixel_decoder_hidden_states = output.pixel_decoder_hidden_states
transformer_decoder_hidden_states = output.transformer_decoder_hidden_states
self.parent.assertTrue(len(encoder_hidden_states), len(config.backbone_config.depths))
self.parent.assertTrue(len(pixel_decoder_hidden_states), len(config.backbone_config.depths))
self.parent.assertTrue(len(transformer_decoder_hidden_states), config.decoder_config.decoder_layers)
def create_and_check_maskformer_model(self, config, pixel_values, pixel_mask, output_hidden_states=False):
with torch.no_grad():
model = MaskFormerModel(config=config)
model.to(torch_device)
model.eval()
output = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
output = model(pixel_values, output_hidden_states=True)
# the correct shape of output.transformer_decoder_hidden_states ensure the correctness of the
# encoder and pixel decoder
self.parent.assertEqual(
output.transformer_decoder_last_hidden_state.shape,
(self.batch_size, self.num_queries, self.mask_feature_size),
)
# let's ensure the other two hidden state exists
self.parent.assertTrue(output.pixel_decoder_last_hidden_state is not None)
self.parent.assertTrue(output.encoder_last_hidden_state is not None)
if output_hidden_states:
self.check_output_hidden_state(output, config)
def create_and_check_maskformer_instance_segmentation_head_model(
self, config, pixel_values, pixel_mask, mask_labels, class_labels
):
model = MaskFormerForInstanceSegmentation(config=config)
model.to(torch_device)
model.eval()
def comm_check_on_output(result):
# let's still check that all the required stuff is there
self.parent.assertTrue(result.transformer_decoder_last_hidden_state is not None)
self.parent.assertTrue(result.pixel_decoder_last_hidden_state is not None)
self.parent.assertTrue(result.encoder_last_hidden_state is not None)
# okay, now we need to check the logits shape
# due to the encoder compression, masks have a //4 spatial size
self.parent.assertEqual(
result.masks_queries_logits.shape,
(self.batch_size, self.num_queries, self.min_size // 4, self.max_size // 4),
)
# + 1 for null class
self.parent.assertEqual(
result.class_queries_logits.shape, (self.batch_size, self.num_queries, self.num_labels + 1)
)
with torch.no_grad():
result = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
result = model(pixel_values)
comm_check_on_output(result)
result = model(
pixel_values=pixel_values, pixel_mask=pixel_mask, mask_labels=mask_labels, class_labels=class_labels
)
comm_check_on_output(result)
self.parent.assertTrue(result.loss is not None)
self.parent.assertEqual(result.loss.shape, torch.Size([]))
@require_torch
class MaskFormerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (MaskFormerModel, MaskFormerForInstanceSegmentation) if is_torch_available() else ()
pipeline_model_mapping = (
{"image-feature-extraction": MaskFormerModel, "image-segmentation": MaskFormerForInstanceSegmentation}
if is_torch_available()
else {}
)
is_encoder_decoder = False
test_pruning = False
test_head_masking = False
test_missing_keys = False
zero_init_hidden_state = True
test_torch_exportable = True
def setUp(self):
self.model_tester = MaskFormerModelTester(self)
self.config_tester = ConfigTester(self, config_class=MaskFormerConfig, has_text_modality=False)
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = copy.deepcopy(inputs_dict)
if return_labels:
if model_class in [MaskFormerForInstanceSegmentation]:
inputs_dict["mask_labels"] = torch.zeros(
(
self.model_tester.batch_size,
self.model_tester.num_labels,
self.model_tester.min_size,
self.model_tester.max_size,
),
dtype=torch.float32,
device=torch_device,
)
inputs_dict["class_labels"] = torch.zeros(
(self.model_tester.batch_size, self.model_tester.num_labels), dtype=torch.long, device=torch_device
)
return inputs_dict
def test_config(self):
self.config_tester.run_common_tests()
def test_maskformer_model(self):
config, inputs = self.model_tester.prepare_config_and_inputs_for_common()
self.model_tester.create_and_check_maskformer_model(config, **inputs, output_hidden_states=False)
def test_maskformer_instance_segmentation_head_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_maskformer_instance_segmentation_head_model(*config_and_inputs)
@unittest.skip(reason="MaskFormer does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="MaskFormer does not have a get_input_embeddings method")
def test_model_get_set_embeddings(self):
pass
@unittest.skip(reason="MaskFormer is not a generative model")
def test_generate_without_input_ids(self):
pass
@unittest.skip(reason="MaskFormer does not use token embeddings")
def test_resize_tokens_embeddings(self):
pass
@require_torch_multi_gpu
@unittest.skip(
reason="MaskFormer has some layers using `add_module` which doesn't work well with `nn.DataParallel`"
)
def test_multi_gpu_data_parallel_forward(self):
pass
@slow
def test_model_from_pretrained(self):
for model_name in ["facebook/maskformer-swin-small-coco"]:
model = MaskFormerModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_model_with_labels(self):
size = (self.model_tester.min_size,) * 2
inputs = {
"pixel_values": torch.randn((2, 3, *size), device=torch_device),
"mask_labels": torch.randn((2, 10, *size), device=torch_device),
"class_labels": torch.zeros(2, 10, device=torch_device).long(),
}
model = MaskFormerForInstanceSegmentation(MaskFormerConfig()).to(torch_device)
outputs = model(**inputs)
self.assertTrue(outputs.loss is not None)
def test_hidden_states_output(self):
config, inputs = self.model_tester.prepare_config_and_inputs_for_common()
self.model_tester.create_and_check_maskformer_model(config, **inputs, output_hidden_states=True)
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._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)
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))
# encoder_hidden_states, pixel_decoder_hidden_states, transformer_decoder_hidden_states, hidden_states
added_hidden_states = 4
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
def test_retain_grad_hidden_states_attentions(self):
# only MaskFormerForInstanceSegmentation has the loss
model_class = self.all_model_classes[1]
config, pixel_values, pixel_mask, mask_labels, class_labels = self.model_tester.prepare_config_and_inputs()
config.output_hidden_states = True
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.train()
outputs = model(pixel_values, mask_labels=mask_labels, class_labels=class_labels)
encoder_hidden_states = outputs.encoder_hidden_states[0]
encoder_hidden_states.retain_grad()
pixel_decoder_hidden_states = outputs.pixel_decoder_hidden_states[0]
pixel_decoder_hidden_states.retain_grad()
# we requires_grad=True in inputs_embeds (line 2152), the original implementation don't
transformer_decoder_hidden_states = outputs.transformer_decoder_hidden_states[0]
transformer_decoder_hidden_states.retain_grad()
attentions = outputs.attentions[0]
attentions.retain_grad()
outputs.loss.backward(retain_graph=True)
self.assertIsNotNone(encoder_hidden_states.grad)
self.assertIsNotNone(pixel_decoder_hidden_states.grad)
self.assertIsNotNone(transformer_decoder_hidden_states.grad)
self.assertIsNotNone(attentions.grad)
def test_forward_auxiliary_loss(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.use_auxiliary_loss = True
config.output_auxiliary_logits = True
config.output_hidden_states = 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_logits)
self.assertEqual(len(outputs.auxiliary_logits), self.model_tester.num_channels - 1)
def test_batching_equivalence(self):
def equivalence(tensor1, tensor2):
return 1.0 - F.cosine_similarity(tensor1.float().flatten(), tensor2.float().flatten(), dim=0, eps=0).max()
def recursive_check(batched_object, single_row_object, model_name, key):
if isinstance(batched_object, (list, tuple)):
for batched_object_value, single_row_object_value in zip(batched_object, single_row_object):
recursive_check(batched_object_value, single_row_object_value, model_name, key)
elif batched_object is None:
return
else:
batched_row = batched_object[:1]
self.assertFalse(
torch.isnan(batched_row).any(), f"Batched output has `nan` in {model_name} for key={key}"
)
self.assertFalse(
torch.isinf(batched_row).any(), f"Batched output has `inf` in {model_name} for key={key}"
)
self.assertFalse(
torch.isnan(single_row_object).any(), f"Single row output has `nan` in {model_name} for key={key}"
)
self.assertFalse(
torch.isinf(single_row_object).any(), f"Single row output has `inf` in {model_name} for key={key}"
)
self.assertTrue(
(equivalence(batched_row, single_row_object)) <= 1e-03,
msg=(
f"Batched and Single row outputs are not equal in {model_name} for key={key}. "
f"Difference={equivalence(batched_row, single_row_object)}."
),
)
config, batched_input = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
config.output_hidden_states = True
model_name = model_class.__name__
batched_input_prepared = self._prepare_for_class(batched_input, model_class)
model = model_class(config).to(torch_device).eval()
batch_size = self.model_tester.batch_size
single_row_input = {}
for key, value in batched_input_prepared.items():
single_batch_shape = value.shape[0] // batch_size
single_row_input[key] = value[:single_batch_shape]
with torch.no_grad():
model_batched_output = model(**batched_input_prepared)
model_row_output = model(**single_row_input)
for key in model_batched_output:
# remove the first zero-init queries to decoder, otherwise cos_similarity = `nan`
# no need to check all hidden_states, already checked separately each one
if key == "transformer_decoder_hidden_states":
model_batched_output[key] = model_batched_output[key][1:]
model_row_output[key] = model_row_output[key][1:]
elif key == "hidden_states":
continue
recursive_check(model_batched_output[key], model_row_output[key], model_name, key)
@require_timm
def test_backbone_selection(self):
config, inputs = self.model_tester.prepare_config_and_inputs_for_common()
config.backbone_config = None
config.backbone_kwargs = {"out_indices": [1, 2, 3]}
config.use_pretrained_backbone = True
# Load a timm backbone
# We can't load transformer checkpoint with timm backbone, as we can't specify features_only and out_indices
config.backbone = "resnet18"
config.use_timm_backbone = True
for model_class in self.all_model_classes:
model = model_class(config).to(torch_device).eval()
if model.__class__.__name__ == "MaskFormerModel":
self.assertEqual(model.pixel_level_module.encoder.out_indices, [1, 2, 3])
elif model.__class__.__name__ == "MaskFormerForUniversalSegmentation":
self.assertEqual(model.model.pixel_level_module.encoder.out_indices, [1, 2, 3])
# Load a HF backbone
config.backbone = "microsoft/resnet-18"
config.use_timm_backbone = False
for model_class in self.all_model_classes:
model = model_class(config).to(torch_device).eval()
if model.__class__.__name__ == "MaskFormerModel":
self.assertEqual(model.pixel_level_module.encoder.out_indices, [1, 2, 3])
elif model.__class__.__name__ == "MaskFormerForUniversalSegmentation":
self.assertEqual(model.model.pixel_level_module.encoder.out_indices, [1, 2, 3])
TOLERANCE = 2e-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_vision
@slow
class MaskFormerModelIntegrationTest(unittest.TestCase):
@cached_property
def default_image_processor(self):
return (
MaskFormerImageProcessor.from_pretrained("facebook/maskformer-swin-small-coco")
if is_vision_available()
else None
)
def test_inference_no_head(self):
model = MaskFormerModel.from_pretrained("facebook/maskformer-swin-small-coco").to(torch_device)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(image, return_tensors="pt").to(torch_device)
inputs_shape = inputs["pixel_values"].shape
# check size is divisible by 32
self.assertTrue((inputs_shape[-1] % 32) == 0 and (inputs_shape[-2] % 32) == 0)
# check size
self.assertEqual(inputs_shape, (1, 3, 800, 1088))
with torch.no_grad():
outputs = model(**inputs)
expected_slice_hidden_state = torch.tensor(
[
[-0.0482, 0.9228, 0.4951],
[-0.2547, 0.8017, 0.8527],
[-0.0069, 0.3385, -0.0089],
]
).to(torch_device)
torch.allclose(outputs.encoder_last_hidden_state[0, 0, :3, :3], expected_slice_hidden_state, atol=TOLERANCE, rtol=TOLERANCE) # fmt: skip
expectations = Expectations(
{
(None, None): [[-0.8422, -0.8434, -0.9718], [-1.0144, -0.5565, -0.4195], [-1.0038, -0.4484, -0.1961]],
("cuda", 8): [
[-0.8422, -0.8435, -0.9717],
[-1.0145, -0.5564, -0.4195],
[-1.0040, -0.4486, -0.1962],
],
}
)
expected_slice_hidden_state = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.allclose(outputs.pixel_decoder_last_hidden_state[0, 0, :3, :3], expected_slice_hidden_state, atol=TOLERANCE,rtol=TOLERANCE) # fmt: skip
expectations = Expectations(
{
(None, None): [
[0.2852, -0.0159, 0.9735],
[0.6254, 0.1858, 0.8529],
[-0.0680, -0.4116, 1.8413],
],
("cuda", 8): [
[0.2853, -0.0162, 0.9736],
[0.6256, 0.1856, 0.8530],
[-0.0679, -0.4118, 1.8416],
],
}
)
expected_slice_hidden_state = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.allclose(outputs.transformer_decoder_last_hidden_state[0, :3, :3], expected_slice_hidden_state, atol=TOLERANCE, rtol=TOLERANCE) # fmt: skip
def test_inference_instance_segmentation_head(self):
model = (
MaskFormerForInstanceSegmentation.from_pretrained("facebook/maskformer-swin-small-coco")
.to(torch_device)
.eval()
)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(image, return_tensors="pt").to(torch_device)
inputs_shape = inputs["pixel_values"].shape
# check size is divisible by 32
self.assertTrue((inputs_shape[-1] % 32) == 0 and (inputs_shape[-2] % 32) == 0)
# check size
self.assertEqual(inputs_shape, (1, 3, 800, 1088))
with torch.no_grad():
outputs = model(**inputs)
# masks_queries_logits
masks_queries_logits = outputs.masks_queries_logits
self.assertEqual(
masks_queries_logits.shape,
(1, model.config.decoder_config.num_queries, inputs_shape[-2] // 4, inputs_shape[-1] // 4),
)
expectations = Expectations(
{
(None, None): [
[-1.3737124, -1.7724937, -1.9364233],
[-1.5977281, -1.9867939, -2.1523695],
[-1.5795398, -1.9269832, -2.093942],
],
("cuda", 8): [
[-1.3737, -1.7727, -1.9367],
[-1.5979, -1.9871, -2.1527],
[-1.5797, -1.9271, -2.0941],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(masks_queries_logits[0, 0, :3, :3], expected_slice, rtol=TOLERANCE, atol=TOLERANCE)
# class_queries_logits
class_queries_logits = outputs.class_queries_logits
self.assertEqual(
class_queries_logits.shape, (1, model.config.decoder_config.num_queries, model.config.num_labels + 1)
)
expectations = Expectations(
{
(None, None): [
[1.6512e00, -5.2572e00, -3.3519e00],
[3.6169e-02, -5.9025e00, -2.9313e00],
[1.0766e-04, -7.7630e00, -5.1263e00],
],
("cuda", 8): [
[1.6512e00, -5.2572e00, -3.3519e00],
[3.6163e-02, -5.9025e00, -2.9313e00],
[1.1681e-04, -7.7631e00, -5.1263e00],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(
outputs.class_queries_logits[0, :3, :3], expected_slice, rtol=TOLERANCE, atol=TOLERANCE
)
def test_inference_instance_segmentation_head_resnet_backbone(self):
model = (
MaskFormerForInstanceSegmentation.from_pretrained("facebook/maskformer-resnet101-coco-stuff")
.to(torch_device)
.eval()
)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(image, return_tensors="pt").to(torch_device)
inputs_shape = inputs["pixel_values"].shape
# check size is divisible by 32
self.assertTrue((inputs_shape[-1] % 32) == 0 and (inputs_shape[-2] % 32) == 0)
# check size
self.assertEqual(inputs_shape, (1, 3, 800, 1088))
with torch.no_grad():
outputs = model(**inputs)
# masks_queries_logits
masks_queries_logits = outputs.masks_queries_logits
self.assertEqual(
masks_queries_logits.shape,
(1, model.config.decoder_config.num_queries, inputs_shape[-2] // 4, inputs_shape[-1] // 4),
)
expectations = Expectations(
{
(None, None): [[-0.9046, -2.6366, -4.6062], [-3.4179, -5.7890, -8.8057], [-4.9179, -7.6560, -10.7711]],
("cuda", 8): [[-0.9046, -2.6366, -4.6062], [-3.4179, -5.7890, -8.8057], [-4.9179, -7.6560, -10.7711]],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(masks_queries_logits[0, 0, :3, :3], expected_slice, rtol=TOLERANCE, atol=TOLERANCE)
# class_queries_logits
class_queries_logits = outputs.class_queries_logits
self.assertEqual(
class_queries_logits.shape, (1, model.config.decoder_config.num_queries, model.config.num_labels + 1)
)
expectations = Expectations(
{
(None, None): [
[4.7188, -3.2585, -2.8857],
[6.6871, -2.9181, -1.2487],
[7.2449, -2.2764, -2.1874],
],
("cuda", 8): [
[4.7188, -3.2585, -2.8857],
[6.6871, -2.9181, -1.2487],
[7.2449, -2.2764, -2.1874],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(
outputs.class_queries_logits[0, :3, :3], expected_slice, rtol=TOLERANCE, atol=TOLERANCE
)
@require_torch_accelerator
@require_torch_fp16
def test_inference_fp16(self):
model = (
MaskFormerForInstanceSegmentation.from_pretrained("facebook/maskformer-resnet101-coco-stuff")
.to(torch_device, dtype=torch.float16)
.eval()
)
image_processor = self.default_image_processor
image = prepare_img()
inputs = image_processor(image, return_tensors="pt").to(torch_device, dtype=torch.float16)
with torch.no_grad():
_ = model(**inputs)
def test_with_segmentation_maps_and_loss(self):
model = (
MaskFormerForInstanceSegmentation.from_pretrained("facebook/maskformer-swin-small-coco")
.to(torch_device)
.eval()
)
image_processor = self.default_image_processor
inputs = image_processor(
[np.zeros((3, 400, 333)), np.zeros((3, 400, 333))],
segmentation_maps=[np.zeros((384, 384)).astype(np.float32), np.zeros((384, 384)).astype(np.float32)],
return_tensors="pt",
)
inputs["pixel_values"] = inputs["pixel_values"].to(torch_device)
inputs["mask_labels"] = [el.to(torch_device) for el in inputs["mask_labels"]]
inputs["class_labels"] = [el.to(torch_device) for el in inputs["class_labels"]]
with torch.no_grad():
outputs = model(**inputs)
self.assertTrue(outputs.loss is not None)

425
transformers/tests/models/maskformer/test_modeling_maskformer_swin.py Normal file
View File

@@ -0,0 +1,425 @@
# 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 MaskFormer Swin model."""
import collections
import unittest
from transformers import MaskFormerSwinConfig
from transformers.testing_utils import require_torch, require_torch_multi_gpu, torch_device
from transformers.utils import is_torch_available
from ...test_backbone_common import BackboneTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from torch import nn
from transformers import MaskFormerSwinBackbone
from transformers.models.maskformer import MaskFormerSwinModel
class MaskFormerSwinModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=32,
patch_size=2,
num_channels=3,
embed_dim=16,
depths=[1, 2, 1],
num_heads=[2, 2, 4],
window_size=2,
mlp_ratio=2.0,
qkv_bias=True,
hidden_dropout_prob=0.0,
attention_probs_dropout_prob=0.0,
drop_path_rate=0.1,
hidden_act="gelu",
use_absolute_embeddings=False,
patch_norm=True,
initializer_range=0.02,
layer_norm_eps=1e-5,
is_training=True,
scope=None,
use_labels=True,
type_sequence_label_size=10,
encoder_stride=8,
out_features=["stage1", "stage2", "stage3"],
out_indices=[1, 2, 3],
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.patch_size = patch_size
self.num_channels = num_channels
self.embed_dim = embed_dim
self.depths = depths
self.num_heads = num_heads
self.window_size = window_size
self.mlp_ratio = mlp_ratio
self.qkv_bias = qkv_bias
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.drop_path_rate = drop_path_rate
self.hidden_act = hidden_act
self.use_absolute_embeddings = use_absolute_embeddings
self.patch_norm = patch_norm
self.layer_norm_eps = layer_norm_eps
self.initializer_range = initializer_range
self.is_training = is_training
self.scope = scope
self.use_labels = use_labels
self.type_sequence_label_size = type_sequence_label_size
self.encoder_stride = encoder_stride
self.out_features = out_features
self.out_indices = out_indices
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
labels = None
if self.use_labels:
labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return MaskFormerSwinConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
embed_dim=self.embed_dim,
depths=self.depths,
num_heads=self.num_heads,
window_size=self.window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=self.qkv_bias,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
drop_path_rate=self.drop_path_rate,
hidden_act=self.hidden_act,
use_absolute_embeddings=self.use_absolute_embeddings,
path_norm=self.patch_norm,
layer_norm_eps=self.layer_norm_eps,
initializer_range=self.initializer_range,
encoder_stride=self.encoder_stride,
out_features=self.out_features,
out_indices=self.out_indices,
)
def create_and_check_model(self, config, pixel_values, labels):
model = MaskFormerSwinModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
expected_seq_len = ((config.image_size // config.patch_size) ** 2) // (4 ** (len(config.depths) - 1))
expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
def create_and_check_backbone(self, config, pixel_values, labels):
model = MaskFormerSwinBackbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
self.parent.assertListEqual(list(result.feature_maps[0].shape), [13, 16, 16, 16])
# verify channels
self.parent.assertEqual(len(model.channels), len(config.out_features))
self.parent.assertListEqual(model.channels, [16, 32, 64])
# verify ValueError
with self.parent.assertRaises(ValueError):
config.out_features = ["stem"]
model = MaskFormerSwinBackbone(config=config)
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 MaskFormerSwinModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
MaskFormerSwinModel,
MaskFormerSwinBackbone,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = {"feature-extraction": MaskFormerSwinModel} if is_torch_available() else {}
fx_compatible = False
test_torchscript = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
test_torch_exportable = True
def setUp(self):
self.model_tester = MaskFormerSwinModelTester(self)
self.config_tester = ConfigTester(
self,
config_class=MaskFormerSwinConfig,
has_text_modality=False,
embed_dim=37,
common_properties=["image_size", "patch_size", "num_channels"],
)
@require_torch_multi_gpu
@unittest.skip(
reason=(
"`MaskFormerSwinModel` outputs `hidden_states_spatial_dimensions` which doesn't work well with"
" `nn.DataParallel`"
)
)
def test_multi_gpu_data_parallel_forward(self):
pass
def test_config(self):
self.config_tester.run_common_tests()
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_backbone(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_backbone(*config_and_inputs)
@unittest.skip(reason="Swin does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Swin does not support feedforward chunking")
def test_feed_forward_chunking(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))
@unittest.skip(reason="MaskFormerSwin is only used as backbone and doesn't support output_attentions")
def test_attention_outputs(self):
pass
def check_hidden_states_output(self, inputs_dict, config, model_class, image_size):
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", len(self.model_tester.depths) + 1
)
self.assertEqual(len(hidden_states), expected_num_layers)
# Swin has a different seq_length
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[num_patches, self.model_tester.embed_dim],
)
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
def test_hidden_states_output_with_padding(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.patch_size = 3
image_size = (
self.model_tester.image_size
if isinstance(self.model_tester.image_size, collections.abc.Iterable)
else (self.model_tester.image_size, self.model_tester.image_size)
)
patch_size = (
config.patch_size
if isinstance(config.patch_size, collections.abc.Iterable)
else (config.patch_size, config.patch_size)
)
padded_height = image_size[0] + patch_size[0] - (image_size[0] % patch_size[0])
padded_width = image_size[1] + patch_size[1] - (image_size[1] % patch_size[1])
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
@unittest.skip(reason="MaskFormerSwin doesn't have pretrained checkpoints")
def test_model_from_pretrained(self):
pass
@unittest.skip(reason="This will be fixed once MaskFormerSwin is replaced by native Swin")
def test_initialization(self):
pass
@unittest.skip(reason="This will be fixed once MaskFormerSwin is replaced by native Swin")
def test_gradient_checkpointing_backward_compatibility(self):
pass
def test_model_outputs_equivalence(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def set_nan_tensor_to_zero(t):
t[t != t] = 0
return t
def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
with torch.no_grad():
tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()
def recursive_check(tuple_object, dict_object):
if isinstance(tuple_object, (list, tuple)):
for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif isinstance(tuple_object, dict):
for tuple_iterable_value, dict_iterable_value in zip(
tuple_object.values(), dict_object.values()
):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif tuple_object is None:
return
else:
self.assertTrue(
torch.allclose(
set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
),
msg=(
"Tuple and dict output are not equal. Difference:"
f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
),
)
recursive_check(tuple_output, dict_output)
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
check_equivalence(model, tuple_inputs, dict_inputs)
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs)
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
@require_torch
class MaskFormerSwinBackboneTest(unittest.TestCase, BackboneTesterMixin):
all_model_classes = (MaskFormerSwinBackbone,) if is_torch_available() else ()
config_class = MaskFormerSwinConfig
def setUp(self):
self.model_tester = MaskFormerSwinModelTester(self)
# Overriding as returned hidden states are tuples of tensors instead of a single tensor
def test_backbone_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
batch_size = inputs_dict["pixel_values"].shape[0]
for backbone_class in self.all_model_classes:
backbone = backbone_class(config)
backbone.to(torch_device)
backbone.eval()
outputs = backbone(**inputs_dict)
# Test default outputs and verify feature maps
self.assertIsInstance(outputs.feature_maps, tuple)
self.assertTrue(len(outputs.feature_maps) == len(backbone.channels))
for feature_map, n_channels in zip(outputs.feature_maps, backbone.channels):
self.assertTrue(feature_map.shape[:2], (batch_size, n_channels))
self.assertIsNone(outputs.hidden_states)
self.assertIsNone(outputs.attentions)
# Test output_hidden_states=True
outputs = backbone(**inputs_dict, output_hidden_states=True)
self.assertIsNotNone(outputs.hidden_states)
self.assertTrue(len(outputs.hidden_states), len(backbone.stage_names))
# We skip the stem layer
for hidden_states, n_channels in zip(outputs.hidden_states[1:], backbone.channels):
for hidden_state in hidden_states:
# Hidden states are in the format (batch_size, (height * width), n_channels)
h_batch_size, _, h_n_channels = hidden_state.shape
self.assertTrue((h_batch_size, h_n_channels), (batch_size, n_channels))
# Test output_attentions=True
if self.has_attentions:
outputs = backbone(**inputs_dict, output_attentions=True)
self.assertIsNotNone(outputs.attentions)