SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Standardize semantic segmentation models outputs (#15469)

Browse Source 
* Standardize instance segmentation models outputs

* Rename output

* Update src/transformers/modeling_outputs.py

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Add legacy argument to the config and model forward

* Update src/transformers/models/beit/modeling_beit.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Copy fix in Segformer

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
This commit is contained in:
Sylvain Gugger
2022-02-04 14:52:07 -05:00
committed by GitHub
parent 31be2f45a9
commit ac6aa10f23
13 changed files with 210 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
tests/test_modeling_beit.py
View File

@@ -92,17 +92,20 @@ class BeitModelTester:
self.initializer_range = initializer_range
self.scope = scope
self.out_indices = out_indices
self.num_labels = num_labels
def prepare_config_and_inputs(self):
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
labels = None
pixel_labels = None
if self.use_labels:
labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
pixel_labels = ids_tensor([self.batch_size, self.image_size, self.image_size], self.num_labels)
config = self.get_config()
return config, pixel_values, labels
return config, pixel_values, labels, pixel_labels
def get_config(self):
return BeitConfig(
@@ -122,7 +125,7 @@ class BeitModelTester:
out_indices=self.out_indices,
)
def create_and_check_model(self, config, pixel_values, labels):
def create_and_check_model(self, config, pixel_values, labels, pixel_labels):
model = BeitModel(config=config)
model.to(torch_device)
model.eval()
@@ -133,7 +136,7 @@ class BeitModelTester:
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
def create_and_check_for_masked_lm(self, config, pixel_values, labels):
def create_and_check_for_masked_lm(self, config, pixel_values, labels, pixel_labels):
model = BeitForMaskedImageModeling(config=config)
model.to(torch_device)
model.eval()
@@ -144,7 +147,7 @@ class BeitModelTester:
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.logits.shape, (self.batch_size, num_patches, self.vocab_size))
def create_and_check_for_image_classification(self, config, pixel_values, labels):
def create_and_check_for_image_classification(self, config, pixel_values, labels, pixel_labels):
config.num_labels = self.type_sequence_label_size
model = BeitForImageClassification(config)
model.to(torch_device)
@@ -152,13 +155,23 @@ class BeitModelTester:
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def create_and_check_for_image_segmentation(self, config, pixel_values, labels, pixel_labels):
config.num_labels = self.num_labels
model = BeitForSemanticSegmentation(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, self.image_size)
)
result = model(pixel_values, labels=pixel_labels)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.num_labels, 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
config, pixel_values, labels, pixel_labels = config_and_inputs
inputs_dict = {"pixel_values": pixel_values}
return config, inputs_dict
@@ -217,6 +230,10 @@ class BeitModelTest(ModelTesterMixin, unittest.TestCase):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*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_training(self):
if not self.model_tester.is_training:
return
@@ -516,14 +533,14 @@ class BeitModelIntegrationTest(unittest.TestCase):
logits = outputs.logits
# verify the logits
expected_shape = torch.Size((1, 150, 160, 160))
expected_shape = torch.Size((1, 150, 640, 640))
self.assertEqual(logits.shape, expected_shape)
expected_slice = torch.tensor(
[
[[-4.9225, -2.3954, -3.0522], [-2.8822, -1.0046, -1.7561], [-2.9549, -1.3228, -2.1347]],
[[-5.8168, -3.4129, -4.0778], [-3.8651, -2.2214, -3.0277], [-3.8356, -2.4643, -3.3535]],
[[-0.0078, 3.9952, 4.0754], [2.9856, 4.6944, 5.0035], [3.2413, 4.7813, 4.9969]],
[[-4.9225, -4.9225, -4.6066], [-4.9225, -4.9225, -4.6066], [-4.6675, -4.6675, -4.3617]],
[[-5.8168, -5.8168, -5.5163], [-5.8168, -5.8168, -5.5163], [-5.5728, -5.5728, -5.2842]],
[[-0.0078, -0.0078, 0.4926], [-0.0078, -0.0078, 0.4926], [0.3664, 0.3664, 0.8309]],
]
).to(torch_device)

29
tests/test_modeling_segformer.py
View File

@@ -133,11 +133,11 @@ class SegformerModelTester:
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.logits.shape, (self.batch_size, self.num_labels, self.image_size, self.image_size)
)
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)
result.logits.shape, (self.batch_size, self.num_labels, self.image_size, self.image_size)
)
def prepare_config_and_inputs_for_common(self):
@@ -245,6 +245,7 @@ class SegformerModelTest(ModelTesterMixin, unittest.TestCase):
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
@@ -255,7 +256,7 @@ class SegformerModelTest(ModelTesterMixin, unittest.TestCase):
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(3, len(outputs))
self.assertEqual(out_len + 1, len(outputs))
self_attentions = outputs.attentions
@@ -357,16 +358,17 @@ class SegformerModelIntegrationTest(unittest.TestCase):
encoded_inputs = feature_extractor(images=image, return_tensors="pt")
pixel_values = encoded_inputs.pixel_values.to(torch_device)
outputs = model(pixel_values)
with torch.no_grad():
outputs = model(pixel_values)
expected_shape = torch.Size((1, model.config.num_labels, 128, 128))
expected_shape = torch.Size((1, model.config.num_labels, 512, 512))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor(
[
[[-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]],
[[-4.6309, -4.6309, -4.7425], [-4.6309, -4.6309, -4.7425], [-4.7011, -4.7011, -4.8136]],
[[-12.1391, -12.1391, -12.2858], [-12.1391, -12.1391, -12.2858], [-12.2309, -12.2309, -12.3758]],
[[-12.5134, -12.5134, -12.6328], [-12.5134, -12.5134, -12.6328], [-12.5576, -12.5576, -12.6865]],
]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3, :3, :3], expected_slice, atol=1e-4))
@@ -385,16 +387,17 @@ class SegformerModelIntegrationTest(unittest.TestCase):
encoded_inputs = feature_extractor(images=image, return_tensors="pt")
pixel_values = encoded_inputs.pixel_values.to(torch_device)
outputs = model(pixel_values)
with torch.no_grad():
outputs = model(pixel_values)
expected_shape = torch.Size((1, model.config.num_labels, 128, 128))
expected_shape = torch.Size((1, model.config.num_labels, 512, 512))
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor(
[
[[-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]],
[[-13.5729, -13.5729, -13.6149], [-13.5729, -13.5729, -13.6149], [-13.6697, -13.6697, -13.7224]],
[[-17.1638, -17.1638, -17.0022], [-17.1638, -17.1638, -17.0022], [-17.1754, -17.1754, -17.0358]],
[[-3.6452, -3.6452, -3.5670], [-3.6452, -3.6452, -3.5670], [-3.5744, -3.5744, -3.5079]],
]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3, :3, :3], expected_slice, atol=1e-1))