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/vitmatte/__init__.py Normal file
View File

361
transformers/tests/models/vitmatte/test_image_processing_vitmatte.py Normal file
View File

@@ -0,0 +1,361 @@
# Copyright 2023 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 time
import unittest
import warnings
import numpy as np
import pytest
from packaging import version
from transformers.image_utils import load_image
from transformers.testing_utils import (
require_torch,
require_torch_accelerator,
require_vision,
slow,
torch_device,
)
from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
from ...test_processing_common import url_to_local_path
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import VitMatteImageProcessor
if is_torchvision_available():
from transformers import VitMatteImageProcessorFast
class VitMatteImageProcessingTester:
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=400,
do_rescale=True,
rescale_factor=0.5,
do_pad=True,
size_divisor=10,
do_normalize=True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
):
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.min_resolution = min_resolution
self.max_resolution = max_resolution
self.do_rescale = do_rescale
self.rescale_factor = rescale_factor
self.do_pad = do_pad
self.size_divisor = size_divisor
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
def prepare_image_processor_dict(self):
return {
"image_mean": self.image_mean,
"image_std": self.image_std,
"do_normalize": self.do_normalize,
"do_rescale": self.do_rescale,
"rescale_factor": self.rescale_factor,
"do_pad": self.do_pad,
"size_divisor": self.size_divisor,
}
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 VitMatteImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = VitMatteImageProcessor if is_vision_available() else None
fast_image_processing_class = VitMatteImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = VitMatteImageProcessingTester(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_rescale"))
self.assertTrue(hasattr(image_processing, "rescale_factor"))
self.assertTrue(hasattr(image_processing, "do_pad"))
self.assertTrue(hasattr(image_processing, "size_divisor"))
# Check size_divisibility for BC, the image proccessor has to have an atribute
self.assertTrue(hasattr(image_processing, "size_divisibility"))
def test_call_numpy(self):
# create random numpy tensors
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
for image in image_inputs:
self.assertIsInstance(image, np.ndarray)
# Test not batched input (image processor does not support batched inputs)
image = image_inputs[0]
trimap = np.random.randint(0, 3, size=image.shape[:2])
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
# Verify that width and height can be divided by size_divisibility and that correct dimensions got merged
self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-3] == 4)
def test_call_pytorch(self):
# 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 not batched input (image processor does not support batched inputs)
image = image_inputs[0]
trimap = np.random.randint(0, 3, size=image.shape[1:])
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
# Verify that width and height can be divided by size_divisibility and that correct dimensions got merged
self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-3] == 4)
# create batched tensors
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True)
image_input = torch.stack(image_inputs, dim=0)
self.assertIsInstance(image_input, torch.Tensor)
self.assertTrue(image_input.shape[1] == 3)
trimap_shape = [image_input.shape[0]] + [1] + list(image_input.shape)[2:]
trimap_input = torch.randint(0, 3, trimap_shape, dtype=torch.uint8)
self.assertIsInstance(trimap_input, torch.Tensor)
self.assertTrue(trimap_input.shape[1] == 1)
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
# Verify that width and height can be divided by size_divisibility and that correct dimensions got merged
self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-3] == 4)
def test_call_pil(self):
# create random PIL images
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
for image in image_inputs:
self.assertIsInstance(image, Image.Image)
# Test not batched input (image processor does not support batched inputs)
image = image_inputs[0]
trimap = np.random.randint(0, 3, size=image.size[::-1])
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
encoded_images = image_processing(images=image, trimaps=trimap, return_tensors="pt").pixel_values
# Verify that width and height can be divided by size_divisibility and that correct dimensions got merged
self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-3] == 4)
def test_call_numpy_4_channels(self):
# Test that can process images which have an arbitrary number of channels
# create random numpy tensors
self.image_processor_tester.num_channels = 4
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
# Test not batched input (image processor does not support batched inputs)
image = image_inputs[0]
trimap = np.random.randint(0, 3, size=image.shape[:2])
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
encoded_images = image_processor(
images=image,
trimaps=trimap,
input_data_format="channels_last",
image_mean=0,
image_std=1,
return_tensors="pt",
).pixel_values
# Verify that width and height can be divided by size_divisibility and that correct dimensions got merged
self.assertTrue(encoded_images.shape[-1] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-2] % self.image_processor_tester.size_divisor == 0)
self.assertTrue(encoded_images.shape[-3] == 5)
def test_padding_slow(self):
image_processing = self.image_processing_class(**self.image_processor_dict)
image = np.random.randn(3, 249, 491)
images = image_processing.pad_image(image)
assert images.shape == (3, 256, 512)
image = np.random.randn(3, 249, 512)
images = image_processing.pad_image(image)
assert images.shape == (3, 256, 512)
def test_padding_fast(self):
# extra test because name is different for fast image processor
image_processing = self.fast_image_processing_class(**self.image_processor_dict)
image = torch.rand(3, 249, 491)
images = image_processing._pad_image(image)
assert images.shape == (3, 256, 512)
image = torch.rand(3, 249, 512)
images = image_processing._pad_image(image)
assert images.shape == (3, 256, 512)
def test_image_processor_preprocess_arguments(self):
# vitmatte require additional trimap input for image_processor
# that is why we override original common test
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
image = self.image_processor_tester.prepare_image_inputs()[0]
trimap = np.random.randint(0, 3, size=image.size[::-1])
with warnings.catch_warnings(record=True) as raised_warnings:
warnings.simplefilter("always")
image_processor(image, trimaps=trimap, extra_argument=True)
messages = " ".join([str(w.message) for w in raised_warnings])
self.assertGreaterEqual(len(raised_warnings), 1)
self.assertIn("extra_argument", messages)
@unittest.skip(reason="Many failing cases. This test needs a more deep investigation.")
def test_fast_is_faster_than_slow(self):
if not self.test_slow_image_processor or not self.test_fast_image_processor:
self.skipTest(reason="Skipping speed test")
if self.image_processing_class is None or self.fast_image_processing_class is None:
self.skipTest(reason="Skipping speed test as one of the image processors is not defined")
def measure_time(image_processor, images, trimaps):
# Warmup
for _ in range(5):
_ = image_processor(images, trimaps=trimaps, return_tensors="pt")
all_times = []
for _ in range(10):
start = time.time()
_ = image_processor(images, trimaps=trimaps, return_tensors="pt")
all_times.append(time.time() - start)
# Take the average of the fastest 3 runs
avg_time = sum(sorted(all_times[:3])) / 3.0
return avg_time
dummy_images = torch.randint(0, 255, (4, 3, 400, 800), dtype=torch.uint8)
dummy_trimaps = torch.randint(0, 3, (4, 400, 800), dtype=torch.uint8)
image_processor_slow = self.image_processing_class(**self.image_processor_dict)
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
fast_time = measure_time(image_processor_fast, dummy_images, dummy_trimaps)
slow_time = measure_time(image_processor_slow, dummy_images, dummy_trimaps)
self.assertLessEqual(fast_time, slow_time)
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 = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
dummy_trimap = np.random.randint(0, 3, size=dummy_image.size[::-1])
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_image, trimaps=dummy_trimap, return_tensors="pt")
encoding_fast = image_processor_fast(dummy_image, trimaps=dummy_trimap, return_tensors="pt")
self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values, encoding_fast.pixel_values)
def test_slow_fast_equivalence_batched(self):
# this only checks on equal resolution, since the slow processor doesn't work otherwise
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 = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True)
dummy_trimaps = [np.random.randint(0, 3, size=image.shape[1:]) for image in dummy_images]
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, trimaps=dummy_trimaps, return_tensors="pt")
encoding_fast = image_processor_fast(dummy_images, trimaps=dummy_trimaps, return_tensors="pt")
self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values, encoding_fast.pixel_values)
@slow
@require_torch_accelerator
@require_vision
@pytest.mark.torch_compile_test
def test_can_compile_fast_image_processor(self):
# override as trimaps are needed for the image processor
if self.fast_image_processing_class is None:
self.skipTest("Skipping compilation test as fast image processor is not defined")
if version.parse(torch.__version__) < version.parse("2.3"):
self.skipTest(reason="This test requires torch >= 2.3 to run.")
torch.compiler.reset()
input_image = torch.randint(0, 255, (3, 224, 224), dtype=torch.uint8)
dummy_trimap = np.random.randint(0, 3, size=input_image.shape[1:])
image_processor = self.fast_image_processing_class(**self.image_processor_dict)
output_eager = image_processor(input_image, dummy_trimap, device=torch_device, return_tensors="pt")
image_processor = torch.compile(image_processor, mode="reduce-overhead")
output_compiled = image_processor(input_image, dummy_trimap, device=torch_device, return_tensors="pt")
torch.testing.assert_close(output_eager.pixel_values, output_compiled.pixel_values, rtol=1e-4, atol=1e-4)

297
transformers/tests/models/vitmatte/test_modeling_vitmatte.py Normal file
View File

@@ -0,0 +1,297 @@
# Copyright 2023 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 VitMatte model."""
import unittest
from huggingface_hub import hf_hub_download
from transformers import VitMatteConfig
from transformers.testing_utils import (
require_timm,
require_torch,
slow,
torch_device,
)
from transformers.utils import is_torch_available, is_vision_available
from transformers.utils.import_utils import get_torch_major_and_minor_version
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 transformers import VitDetConfig, VitMatteForImageMatting
if is_vision_available():
from PIL import Image
from transformers import VitMatteImageProcessor
class VitMatteModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=32,
patch_size=16,
num_channels=4,
is_training=True,
use_labels=False,
hidden_size=2,
num_hidden_layers=2,
num_attention_heads=2,
hidden_act="gelu",
type_sequence_label_size=10,
initializer_range=0.02,
scope=None,
out_features=["stage1"],
fusion_hidden_sizes=[128, 64, 32, 16],
):
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.hidden_act = hidden_act
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.scope = scope
self.out_features = out_features
self.fusion_hidden_sizes = fusion_hidden_sizes
self.seq_length = (self.image_size // self.patch_size) ** 2
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:
raise NotImplementedError("Training is not yet supported")
config = self.get_config()
return config, pixel_values, labels
def get_backbone_config(self):
return VitDetConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
hidden_size=self.hidden_size,
is_training=self.is_training,
hidden_act=self.hidden_act,
out_features=self.out_features,
)
def get_config(self):
return VitMatteConfig(
backbone_config=self.get_backbone_config(),
backbone=None,
hidden_size=self.hidden_size,
fusion_hidden_sizes=self.fusion_hidden_sizes,
)
def create_and_check_model(self, config, pixel_values, labels):
model = VitMatteForImageMatting(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(result.alphas.shape, (self.batch_size, 1, self.image_size, self.image_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@require_torch
class VitMatteModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
"""
Here we also overwrite some of the tests of test_modeling_common.py, as VitMatte does not use input_ids, inputs_embeds,
attention_mask and seq_length.
"""
all_model_classes = (VitMatteForImageMatting,) if is_torch_available() else ()
pipeline_model_mapping = {}
fx_compatible = False
test_pruning = False
test_resize_embeddings = False
test_head_masking = False
test_torch_exportable = True
test_torch_exportable_strictly = get_torch_major_and_minor_version() != "2.7"
def setUp(self):
self.model_tester = VitMatteModelTester(self)
self.config_tester = ConfigTester(
self,
config_class=VitMatteConfig,
has_text_modality=False,
hidden_size=37,
common_properties=["hidden_size"],
)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="VitMatte does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training(self):
pass
@unittest.skip(reason="Training is not yet supported")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="ViTMatte does not support input and output embeddings")
def test_model_get_set_embeddings(self):
pass
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
model_name = "hustvl/vitmatte-small-composition-1k"
model = VitMatteForImageMatting.from_pretrained(model_name)
self.assertIsNotNone(model)
@unittest.skip(reason="ViTMatte does not support retaining gradient on attention logits")
def test_retain_grad_hidden_states_attentions(self):
pass
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.encoder_hidden_states if config.is_encoder_decoder else 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)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[2, 2],
)
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
print("Hello we're here")
check_hidden_states_output(inputs_dict, config, model_class)
@require_timm
def test_backbone_selection(self):
def _validate_backbone_init():
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
if model.__class__.__name__ == "VitMatteForImageMatting":
# Confirm out_indices propagated to backbone
self.assertEqual(len(model.backbone.out_indices), 2)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.use_pretrained_backbone = True
config.backbone_config = None
config.backbone_kwargs = {"out_indices": [-2, -1]}
# Force load_backbone path
config.is_hybrid = False
# Load a timm backbone
config.backbone = "resnet18"
config.use_timm_backbone = True
_validate_backbone_init()
# Load a HF backbone
config.backbone = "facebook/dinov2-small"
config.use_timm_backbone = False
_validate_backbone_init()
@require_torch
class VitMatteModelIntegrationTest(unittest.TestCase):
@slow
def test_inference(self):
processor = VitMatteImageProcessor.from_pretrained("hustvl/vitmatte-small-composition-1k")
model = VitMatteForImageMatting.from_pretrained("hustvl/vitmatte-small-composition-1k").to(torch_device)
filepath = hf_hub_download(
repo_id="hf-internal-testing/image-matting-fixtures", filename="image.png", repo_type="dataset"
)
image = Image.open(filepath).convert("RGB")
filepath = hf_hub_download(
repo_id="hf-internal-testing/image-matting-fixtures", filename="trimap.png", repo_type="dataset"
)
trimap = Image.open(filepath).convert("L")
# prepare image + trimap for the model
inputs = processor(images=image, trimaps=trimap, return_tensors="pt").to(torch_device)
with torch.no_grad():
alphas = model(**inputs).alphas
expected_shape = torch.Size((1, 1, 640, 960))
self.assertEqual(alphas.shape, expected_shape)
expected_slice = torch.tensor(
[[0.9977, 0.9987, 0.9990], [0.9980, 0.9998, 0.9998], [0.9983, 0.9998, 0.9998]], device=torch_device
)
torch.testing.assert_close(alphas[0, 0, :3, :3], expected_slice, rtol=1e-4, atol=1e-4)