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

Backbone add mixin tests (#22542)

Browse Source 
* Add out_indices to backbones, deprecate out_features

* Update - can specify both out_features and out_indices but not both

* Add backbone mixin tests

* Test tidy up

* Add test_backbone for convnext

* Remove redefinition of method

* Update for Dinat and Nat backbones

* Update tests

* Smarter indexing

* Add checks on config creation for backbone

* PR comments
This commit is contained in:
amyeroberts
2023-04-06 13:50:15 +01:00
committed by GitHub
parent 48706c7178
commit 12d51db243
10 changed files with 316 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
tests/models/bit/test_modeling_bit.py
View File

@@ -22,6 +22,7 @@ from transformers import BitConfig
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_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
@@ -55,6 +56,7 @@ class BitModelTester:
num_labels=3,
scope=None,
out_features=["stage2", "stage3", "stage4"],
out_indices=[2, 3, 4],
num_groups=1,
):
self.parent = parent
@@ -71,6 +73,7 @@ class BitModelTester:
self.scope = scope
self.num_stages = len(hidden_sizes)
self.out_features = out_features
self.out_indices = out_indices
self.num_groups = num_groups
def prepare_config_and_inputs(self):
@@ -93,6 +96,7 @@ class BitModelTester:
hidden_act=self.hidden_act,
num_labels=self.num_labels,
out_features=self.out_features,
out_indices=self.out_indices,
num_groups=self.num_groups,
)
@@ -317,3 +321,14 @@ class BitModelIntegrationTest(unittest.TestCase):
expected_slice = torch.tensor([[-0.6526, -0.5263, -1.4398]]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
class BitBackboneTest(BackboneTesterMixin, unittest.TestCase):
all_model_classes = (BitBackbone,) if is_torch_available() else ()
config_class = BitConfig
has_attentions = False
def setUp(self):
self.model_tester = BitModelTester(self)

19
tests/models/convnext/test_modeling_convnext.py
View File

@@ -22,6 +22,7 @@ from transformers import ConvNextConfig
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_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
@@ -57,6 +58,7 @@ class ConvNextModelTester:
num_labels=10,
initializer_range=0.02,
out_features=["stage2", "stage3", "stage4"],
out_indices=[2, 3, 4],
scope=None,
):
self.parent = parent
@@ -73,6 +75,7 @@ class ConvNextModelTester:
self.num_labels = num_labels
self.initializer_range = initializer_range
self.out_features = out_features
self.out_indices = out_indices
self.scope = scope
def prepare_config_and_inputs(self):
@@ -95,6 +98,7 @@ class ConvNextModelTester:
is_decoder=False,
initializer_range=self.initializer_range,
out_features=self.out_features,
out_indices=self.out_indices,
num_labels=self.num_labels,
)
@@ -224,6 +228,10 @@ class ConvNextModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
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)
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
@@ -299,3 +307,14 @@ class ConvNextModelIntegrationTest(unittest.TestCase):
expected_slice = torch.tensor([-0.0260, -0.4739, 0.1911]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
class ConvNextBackboneTest(unittest.TestCase, BackboneTesterMixin):
all_model_classes = (ConvNextBackbone,) if is_torch_available() else ()
config_class = ConvNextConfig
has_attentions = False
def setUp(self):
self.model_tester = ConvNextModelTester(self)

3
tests/models/convnextv2/test_modeling_convnextv2.py
View File

@@ -58,6 +58,7 @@ class ConvNextV2ModelTester:
num_labels=10,
initializer_range=0.02,
out_features=["stage2", "stage3", "stage4"],
out_indices=[2, 3, 4],
scope=None,
):
self.parent = parent
@@ -74,6 +75,7 @@ class ConvNextV2ModelTester:
self.num_labels = num_labels
self.initializer_range = initializer_range
self.out_features = out_features
self.out_indices = out_indices
self.scope = scope
def prepare_config_and_inputs(self):
@@ -97,6 +99,7 @@ class ConvNextV2ModelTester:
is_decoder=False,
initializer_range=self.initializer_range,
out_features=self.out_features,
out_indices=self.out_indices,
num_labels=self.num_labels,
)

14
tests/models/dinat/test_modeling_dinat.py
View File

@@ -22,6 +22,7 @@ from transformers import DinatConfig
from transformers.testing_utils import require_natten, require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_available
from ...test_backbone_common import BackboneTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
@@ -67,6 +68,7 @@ class DinatModelTester:
use_labels=True,
num_labels=10,
out_features=["stage1", "stage2"],
out_indices=[1, 2],
):
self.parent = parent
self.batch_size = batch_size
@@ -92,6 +94,7 @@ class DinatModelTester:
self.use_labels = use_labels
self.num_labels = num_labels
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])
@@ -125,6 +128,7 @@ class DinatModelTester:
layer_norm_eps=self.layer_norm_eps,
initializer_range=self.initializer_range,
out_features=self.out_features,
out_indices=self.out_indices,
)
def create_and_check_model(self, config, pixel_values, labels):
@@ -383,3 +387,13 @@ class DinatModelIntegrationTest(unittest.TestCase):
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor([-0.1545, -0.7667, 0.4642]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
@require_natten
class DinatBackboneTest(unittest.TestCase, BackboneTesterMixin):
all_model_classes = (DinatBackbone,) if is_torch_available() else ()
config_class = DinatConfig
def setUp(self):
self.model_tester = DinatModelTester(self)

49
tests/models/maskformer/test_modeling_maskformer_swin.py
View File

@@ -23,6 +23,7 @@ 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
@@ -64,6 +65,7 @@ class MaskFormerSwinModelTester:
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
@@ -90,6 +92,7 @@ class MaskFormerSwinModelTester:
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])
@@ -123,6 +126,7 @@ class MaskFormerSwinModelTester:
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):
@@ -395,3 +399,48 @@ class MaskFormerSwinModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.Te
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)

14
tests/models/nat/test_modeling_nat.py
View File

@@ -22,6 +22,7 @@ from transformers import NatConfig
from transformers.testing_utils import require_natten, require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_available
from ...test_backbone_common import BackboneTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
@@ -66,6 +67,7 @@ class NatModelTester:
use_labels=True,
num_labels=10,
out_features=["stage1", "stage2"],
out_indices=[1, 2],
):
self.parent = parent
self.batch_size = batch_size
@@ -90,6 +92,7 @@ class NatModelTester:
self.use_labels = use_labels
self.num_labels = num_labels
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])
@@ -122,6 +125,7 @@ class NatModelTester:
layer_norm_eps=self.layer_norm_eps,
initializer_range=self.initializer_range,
out_features=self.out_features,
out_indices=self.out_indices,
)
def create_and_check_model(self, config, pixel_values, labels):
@@ -380,3 +384,13 @@ class NatModelIntegrationTest(unittest.TestCase):
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor([0.3805, -0.8676, -0.3912]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
@require_natten
class NatBackboneTest(unittest.TestCase, BackboneTesterMixin):
all_model_classes = (NatBackbone,) if is_torch_available() else ()
config_class = NatConfig
def setUp(self):
self.model_tester = NatModelTester(self)

14
tests/models/resnet/test_modeling_resnet.py
View File

@@ -22,6 +22,7 @@ from transformers import ResNetConfig
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_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
@@ -57,6 +58,7 @@ class ResNetModelTester:
num_labels=3,
scope=None,
out_features=["stage2", "stage3", "stage4"],
out_indices=[2, 3, 4],
):
self.parent = parent
self.batch_size = batch_size
@@ -72,6 +74,7 @@ class ResNetModelTester:
self.scope = scope
self.num_stages = len(hidden_sizes)
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])
@@ -93,6 +96,7 @@ class ResNetModelTester:
hidden_act=self.hidden_act,
num_labels=self.num_labels,
out_features=self.out_features,
out_indices=self.out_indices,
)
def create_and_check_model(self, config, pixel_values, labels):
@@ -323,3 +327,13 @@ class ResNetModelIntegrationTest(unittest.TestCase):
expected_slice = torch.tensor([-11.1069, -9.7877, -8.3777]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
class ResNetBackboneTest(BackboneTesterMixin, unittest.TestCase):
all_model_classes = (ResNetBackbone,) if is_torch_available() else ()
has_attentions = False
config_class = ResNetConfig
def setUp(self):
self.model_tester = ResNetModelTester(self)

13
tests/models/swin/test_modeling_swin.py
View File

@@ -22,6 +22,7 @@ from transformers import SwinConfig
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
from transformers.utils import cached_property, is_torch_available, is_vision_available
from ...test_backbone_common import BackboneTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
@@ -69,6 +70,7 @@ class SwinModelTester:
type_sequence_label_size=10,
encoder_stride=8,
out_features=["stage1", "stage2"],
out_indices=[1, 2],
):
self.parent = parent
self.batch_size = batch_size
@@ -95,6 +97,7 @@ class SwinModelTester:
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])
@@ -128,6 +131,7 @@ class SwinModelTester:
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):
@@ -502,3 +506,12 @@ class SwinModelIntegrationTest(unittest.TestCase):
self.assertEqual(outputs.logits.shape, expected_shape)
expected_slice = torch.tensor([-0.0948, -0.6454, -0.0921]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@require_torch
class SwinBackboneTest(unittest.TestCase, BackboneTesterMixin):
all_model_classes = (SwinBackbone,) if is_torch_available() else ()
config_class = SwinConfig
def setUp(self):
self.model_tester = SwinModelTester(self)