EngineX-Cambricon/enginex-mlu370-any2any
10
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/segformer/__init__.py Normal file
View File

302
transformers/tests/models/segformer/test_image_processing_segformer.py Normal file
View File

@@ -0,0 +1,302 @@
# 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 unittest
from datasets import load_dataset
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 SegformerImageProcessor
if is_torchvision_available():
from transformers import SegformerImageProcessorFast
class SegformerImageProcessingTester:
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_reduce_labels=False,
):
size = size if size is not None else {"height": 30, "width": 30}
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_reduce_labels = do_reduce_labels
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_reduce_labels": self.do_reduce_labels,
}
def expected_output_image_shape(self, images):
return self.num_channels, self.size["height"], self.size["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,
)
def prepare_semantic_single_inputs():
ds = load_dataset("hf-internal-testing/fixtures_ade20k", split="test")
example = ds[0]
return example["image"], example["map"]
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 SegformerImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = SegformerImageProcessor if is_vision_available() else None
fast_image_processing_class = SegformerImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = SegformerImageProcessingTester(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, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
self.assertTrue(hasattr(image_processing, "do_reduce_labels"))
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, {"height": 30, "width": 30})
self.assertEqual(image_processor.do_reduce_labels, False)
image_processor = image_processing_class.from_dict(
self.image_processor_dict, size=42, do_reduce_labels=True
)
self.assertEqual(image_processor.size, {"height": 42, "width": 42})
self.assertEqual(image_processor.do_reduce_labels, True)
def test_call_segmentation_maps(self):
for image_processing_class in self.image_processor_list:
# Initialize image_processing
image_processing = image_processing_class(**self.image_processor_dict)
# create random PyTorch tensors
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
maps = []
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
maps.append(torch.zeros(image.shape[-2:]).long())
# Test not batched input
encoding = image_processing(image_inputs[0], maps[0], return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
1,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
1,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Test batched
encoding = image_processing(image_inputs, maps, return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
self.image_processor_tester.batch_size,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Test not batched input (PIL images)
image, segmentation_map = prepare_semantic_single_inputs()
encoding = image_processing(image, segmentation_map, return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
1,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
1,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
# Test batched input (PIL images)
images, segmentation_maps = prepare_semantic_batch_inputs()
encoding = image_processing(images, segmentation_maps, return_tensors="pt")
self.assertEqual(
encoding["pixel_values"].shape,
(
2,
self.image_processor_tester.num_channels,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(
encoding["labels"].shape,
(
2,
self.image_processor_tester.size["height"],
self.image_processor_tester.size["width"],
),
)
self.assertEqual(encoding["labels"].dtype, torch.long)
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
def test_reduce_labels(self):
# Initialize image_processing
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
# ADE20k has 150 classes, and the background is included, so labels should be between 0 and 150
image, map = prepare_semantic_single_inputs()
encoding = image_processing(image, map, return_tensors="pt")
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 150)
image_processing.do_reduce_labels = True
encoding = image_processing(image, map, return_tensors="pt")
self.assertTrue(encoding["labels"].min().item() >= 0)
self.assertTrue(encoding["labels"].max().item() <= 255)
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)
self._assert_slow_fast_tensors_equivalence(
image_encoding_slow.labels.float(), image_encoding_fast.labels.float(), atol=5, mean_atol=0.01
)
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)
self._assert_slow_fast_tensors_equivalence(
encoding_slow.labels.float(), encoding_fast.labels.float(), atol=5, mean_atol=0.01
)

455
transformers/tests/models/segformer/test_modeling_segformer.py Normal file
View File

@@ -0,0 +1,455 @@
# Copyright 2021 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 SegFormer model."""
import unittest
from transformers import SegformerConfig, is_torch_available, is_vision_available
from transformers.testing_utils import Expectations, require_torch, slow, torch_device
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 transformers import (
SegformerForImageClassification,
SegformerForSemanticSegmentation,
SegformerModel,
)
from transformers.models.auto.modeling_auto import MODEL_MAPPING_NAMES
if is_vision_available():
from PIL import Image
from transformers import SegformerImageProcessor
class SegformerConfigTester(ConfigTester):
def create_and_test_config_common_properties(self):
config = self.config_class(**self.inputs_dict)
self.parent.assertTrue(hasattr(config, "hidden_sizes"))
self.parent.assertTrue(hasattr(config, "num_attention_heads"))
self.parent.assertTrue(hasattr(config, "num_encoder_blocks"))
class SegformerModelTester:
def __init__(
self,
parent,
batch_size=13,
image_size=64,
num_channels=3,
num_encoder_blocks=4,
depths=[1, 1, 1, 1],
sr_ratios=[8, 4, 2, 1],
hidden_sizes=[8, 8, 16, 16],
downsampling_rates=[1, 4, 8, 16],
num_attention_heads=[1, 1, 2, 2],
is_training=True,
use_labels=True,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
initializer_range=0.02,
num_labels=3,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.image_size = image_size
self.num_channels = num_channels
self.num_encoder_blocks = num_encoder_blocks
self.sr_ratios = sr_ratios
self.depths = depths
self.hidden_sizes = hidden_sizes
self.downsampling_rates = downsampling_rates
self.num_attention_heads = num_attention_heads
self.is_training = is_training
self.use_labels = use_labels
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.initializer_range = initializer_range
self.num_labels = num_labels
self.scope = scope
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.image_size, self.image_size], self.num_labels)
config = self.get_config()
return config, pixel_values, labels
def get_config(self):
return SegformerConfig(
image_size=self.image_size,
num_channels=self.num_channels,
num_encoder_blocks=self.num_encoder_blocks,
depths=self.depths,
hidden_sizes=self.hidden_sizes,
num_attention_heads=self.num_attention_heads,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, pixel_values, labels):
model = SegformerModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
expected_height = expected_width = self.image_size // (self.downsampling_rates[-1] * 2)
self.parent.assertEqual(
result.last_hidden_state.shape, (self.batch_size, self.hidden_sizes[-1], expected_height, expected_width)
)
def create_and_check_for_image_segmentation(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = SegformerForSemanticSegmentation(config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_labels, self.image_size // 4, self.image_size // 4)
)
result = model(pixel_values, labels=labels)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_labels, self.image_size // 4, self.image_size // 4)
)
self.parent.assertGreater(result.loss, 0.0)
def create_and_check_for_binary_image_segmentation(self, config, pixel_values, labels):
config.num_labels = 1
model = SegformerForSemanticSegmentation(config=config)
model.to(torch_device)
model.eval()
labels = torch.randint(0, 1, (self.batch_size, self.image_size, self.image_size)).to(torch_device)
result = model(pixel_values, labels=labels)
self.parent.assertGreater(result.loss, 0.0)
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 SegformerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
SegformerModel,
SegformerForSemanticSegmentation,
SegformerForImageClassification,
)
if is_torch_available()
else ()
)
pipeline_model_mapping = (
{
"image-feature-extraction": SegformerModel,
"image-classification": SegformerForImageClassification,
"image-segmentation": SegformerForSemanticSegmentation,
}
if is_torch_available()
else {}
)
fx_compatible = True
test_head_masking = False
test_pruning = False
test_resize_embeddings = False
test_torch_exportable = True
def setUp(self):
self.model_tester = SegformerModelTester(self)
self.config_tester = SegformerConfigTester(self, config_class=SegformerConfig)
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_for_binary_image_segmentation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_binary_image_segmentation(*config_and_inputs)
def test_for_image_segmentation(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_image_segmentation(*config_and_inputs)
def test_batching_equivalence(self, atol=2e-4, rtol=2e-4):
super().test_batching_equivalence(atol=atol, rtol=rtol)
@unittest.skip(reason="SegFormer does not use inputs_embeds")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="SegFormer does not have get_input_embeddings method and get_output_embeddings methods")
def test_model_get_set_embeddings(self):
pass
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class._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
expected_num_attentions = sum(self.model_tester.depths)
self.assertEqual(len(attentions), expected_num_attentions)
# 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), expected_num_attentions)
# verify the first attentions (first block, first layer)
expected_seq_len = (self.model_tester.image_size // 4) ** 2
expected_reduced_seq_len = (self.model_tester.image_size // (4 * self.model_tester.sr_ratios[0])) ** 2
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads[0], expected_seq_len, expected_reduced_seq_len],
)
# verify the last attentions (last block, last layer)
expected_seq_len = (self.model_tester.image_size // 32) ** 2
expected_reduced_seq_len = (self.model_tester.image_size // (32 * self.model_tester.sr_ratios[-1])) ** 2
self.assertListEqual(
list(attentions[-1].shape[-3:]),
[self.model_tester.num_attention_heads[-1], expected_seq_len, expected_reduced_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))
self.assertEqual(out_len + 1, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), expected_num_attentions)
# verify the first attentions (first block, first layer)
expected_seq_len = (self.model_tester.image_size // 4) ** 2
expected_reduced_seq_len = (self.model_tester.image_size // (4 * self.model_tester.sr_ratios[0])) ** 2
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads[0], expected_seq_len, expected_reduced_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 = self.model_tester.num_encoder_blocks
self.assertEqual(len(hidden_states), expected_num_layers)
# verify the first hidden states (first block)
self.assertListEqual(
list(hidden_states[0].shape[-3:]),
[
self.model_tester.hidden_sizes[0],
self.model_tester.image_size // 4,
self.model_tester.image_size // 4,
],
)
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_training(self):
if not self.model_tester.is_training:
self.skipTest(reason="model_tester.is_training is set to False")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
for model_class in self.all_model_classes:
if model_class.__name__ in MODEL_MAPPING_NAMES.values():
continue
model = model_class(config)
model.to(torch_device)
model.train()
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
loss = model(**inputs).loss
loss.backward()
@slow
def test_model_from_pretrained(self):
model_name = "nvidia/segformer-b0-finetuned-ade-512-512"
model = SegformerModel.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
class SegformerModelIntegrationTest(unittest.TestCase):
@slow
def test_inference_image_segmentation_ade(self):
# only resize + normalize
image_processor = SegformerImageProcessor(
image_scale=(512, 512), keep_ratio=False, align=False, do_random_crop=False
)
model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512").to(
torch_device
)
image = prepare_img()
encoded_inputs = image_processor(images=image, return_tensors="pt")
pixel_values = encoded_inputs.pixel_values.to(torch_device)
with torch.no_grad():
outputs = model(pixel_values)
expected_shape = torch.Size((1, model.config.num_labels, 128, 128))
self.assertEqual(outputs.logits.shape, expected_shape)
expectations = Expectations(
{
(None, None): [
[[-4.6310, -5.5232, -6.2356], [-5.1921, -6.1444, -6.5996], [-5.4424, -6.2790, -6.7574]],
[[-12.1391, -13.3122, -13.9554], [-12.8732, -13.9352, -14.3563], [-12.9438, -13.8226, -14.2513]],
[[-12.5134, -13.4686, -14.4915], [-12.8669, -14.4343, -14.7758], [-13.2523, -14.5819, -15.0694]],
],
("cuda", 8): [
[[-4.6310, -5.5232, -6.2361], [-5.1918, -6.1445, -6.5996], [-5.4427, -6.2792, -6.7580]],
[[-12.1397, -13.3124, -13.9551], [-12.8736, -13.9347, -14.3569], [-12.9440, -13.8222, -14.2514]],
[[-12.5135, -13.4682, -14.4913], [-12.8670, -14.4339, -14.7766], [-13.2519, -14.5800, -15.0685]],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(outputs.logits[0, :3, :3, :3], expected_slice, rtol=2e-4, atol=2e-4)
@slow
def test_inference_image_segmentation_city(self):
# only resize + normalize
image_processor = SegformerImageProcessor(
image_scale=(512, 512), keep_ratio=False, align=False, do_random_crop=False
)
model = SegformerForSemanticSegmentation.from_pretrained(
"nvidia/segformer-b1-finetuned-cityscapes-1024-1024"
).to(torch_device)
image = prepare_img()
encoded_inputs = image_processor(images=image, return_tensors="pt")
pixel_values = encoded_inputs.pixel_values.to(torch_device)
with torch.no_grad():
outputs = model(pixel_values)
expected_shape = torch.Size((1, model.config.num_labels, 128, 128))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor([]).to(torch_device)
expectations = Expectations(
{
(None, None): [
[[-13.5748, -13.9111, -12.6500], [-14.3500, -15.3683, -14.2328], [-14.7532, -16.0424, -15.6087]],
[[-17.1651, -15.8725, -12.9653], [-17.2580, -17.3718, -14.8223], [-16.6058, -16.8783, -16.7452]],
[[-3.6456, -3.0209, -1.4203], [-3.0797, -3.1959, -2.0000], [-1.8757, -1.9217, -1.6997]],
],
("cuda", 8): [
[[-13.5728, -13.9089, -12.6492], [-14.3478, -15.3656, -14.2309], [-14.7512, -16.0394, -15.6065]],
[[-17.1642, -15.8704, -12.9641], [-17.2572, -17.3701, -14.8214], [-16.6043, -16.8761, -16.7425]],
[[-3.6444, -3.0189, -1.4195], [-3.0787, -3.1953, -1.9993], [-1.8755, -1.9219, -1.7002]],
],
}
)
expected_slice = torch.tensor(expectations.get_expectation()).to(torch_device)
torch.testing.assert_close(outputs.logits[0, :3, :3, :3], expected_slice, rtol=1e-1, atol=1e-1)
@slow
def test_post_processing_semantic_segmentation(self):
# only resize + normalize
image_processor = SegformerImageProcessor(
image_scale=(512, 512), keep_ratio=False, align=False, do_random_crop=False
)
model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512").to(
torch_device
)
image = prepare_img()
encoded_inputs = image_processor(images=image, return_tensors="pt")
pixel_values = encoded_inputs.pixel_values.to(torch_device)
with torch.no_grad():
outputs = model(pixel_values)
outputs.logits = outputs.logits.detach().cpu()
segmentation = image_processor.post_process_semantic_segmentation(outputs=outputs, target_sizes=[(500, 300)])
expected_shape = torch.Size((500, 300))
self.assertEqual(segmentation[0].shape, expected_shape)
segmentation = image_processor.post_process_semantic_segmentation(outputs=outputs)
expected_shape = torch.Size((128, 128))
self.assertEqual(segmentation[0].shape, expected_shape)