Use our data format converter in model tests

test/models/test_alexnet.py

@@ -1,17 +1,12 @@
 import numpy as np
-import pytest
-import tensorflow as tf
+from torchvision.models import alexnet
 
 from test.utils import convert_and_test
-from torchvision.models import alexnet
 
 
-@pytest.mark.parametrize('change_ordering', [True, False])
-def test_alexnet(change_ordering):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_alexnet():
     model = alexnet()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_deeplab.py

@@ -6,12 +6,10 @@
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [False])
 @pytest.mark.parametrize('model_class', [deeplabv3_resnet50, deeplabv3_resnet101, deeplabv3_mobilenet_v3_large])
-def test_deeplab(change_ordering, model_class):
+def test_deeplab(model_class):
     model = model_class()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 256, 256))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering,
-                             should_transform_inputs=True)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_densenet.py

@@ -1,18 +1,14 @@
 import numpy as np
 import pytest
-import tensorflow as tf
+from torchvision.models.densenet import densenet121
 
 from test.utils import convert_and_test
-from torchvision.models.densenet import densenet121
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [True, False])
-def test_densenet(change_ordering):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_densenet():
     model = densenet121()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_googlenet.py

@@ -1,18 +1,14 @@
 import numpy as np
 import pytest
-import tensorflow as tf
+from torchvision.models import googlenet
 
 from test.utils import convert_and_test
-from torchvision.models import googlenet
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [True, False])
-def test_googlenet(change_ordering):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_googlenet():
     model = googlenet()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_mbnet2.py

@@ -1,17 +1,12 @@
 import numpy as np
-import pytest
-import tensorflow as tf
+from torchvision.models import mobilenet_v2
 
 from test.utils import convert_and_test
-from torchvision.models import mobilenet_v2
 
 
-@pytest.mark.parametrize('change_ordering', [True, False])
-def test_mobilenetv2(change_ordering):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_mobilenetv2():
     model = mobilenet_v2()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_mnasnet.py

@@ -1,19 +1,15 @@
 import numpy as np
 import pytest
-import tensorflow as tf
+from torchvision.models import mnasnet0_5, mnasnet1_0, mnasnet0_75, mnasnet1_3
 
 from test.utils import convert_and_test
-from torchvision.models import mnasnet0_5, mnasnet1_0, mnasnet0_75, mnasnet1_3
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [True, False])
 @pytest.mark.parametrize('model_class', [mnasnet0_5, mnasnet0_75, mnasnet1_0, mnasnet1_3])
-def test_mnasnet(change_ordering, model_class):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_mnasnet(model_class):
     model = model_class()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_resnet18.py

@@ -1,18 +1,14 @@
 import numpy as np
 import pytest
-import tensorflow as tf
-
 from torchvision.models import resnet18
 
 from test.utils import convert_and_test
 
 
-@pytest.mark.parametrize('change_ordering', [True, False])
+@pytest.mark.parametrize('change_ordering', [False])
 def test_resnet18(change_ordering):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
     model = resnet18()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_resnext.py

@@ -1,19 +1,15 @@
 import numpy as np
 import pytest
-import tensorflow as tf
+from torchvision.models import resnext50_32x4d, resnext101_32x8d
 
 from test.utils import convert_and_test
-from torchvision.models import resnext50_32x4d, resnext101_32x8d
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [True, False])
 @pytest.mark.parametrize('model_class', [resnext50_32x4d, resnext101_32x8d])
-def test_resnext(change_ordering, model_class):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_resnext(model_class):
     model = model_class()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_squeezenet.py

@@ -1,19 +1,15 @@
 import numpy as np
 import pytest
-import tensorflow as tf
+from torchvision.models import squeezenet1_0, squeezenet1_1
 
 from test.utils import convert_and_test
-from torchvision.models import squeezenet1_0, squeezenet1_1
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [True, False])
 @pytest.mark.parametrize('model_class', [squeezenet1_1, squeezenet1_0])
-def test_squeezenet(change_ordering, model_class):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_squeezenet(model_class):
     model = model_class()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/models/test_vgg.py

@@ -1,19 +1,15 @@
 import numpy as np
 import pytest
-import tensorflow as tf
+from torchvision.models import vgg11, vgg11_bn
 
 from test.utils import convert_and_test
-from torchvision.models import vgg11, vgg11_bn
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize('change_ordering', [True, False])
 @pytest.mark.parametrize('model_class', [vgg11, vgg11_bn])
-def test_vgg(change_ordering, model_class):
-    if not tf.test.gpu_device_name() and not change_ordering:
-        pytest.skip("Skip! Since tensorflow Conv2D op currently only supports the NHWC tensor format on the CPU")
+def test_vgg(model_class):
     model = model_class()
     model.eval()
 
     input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
-    error = convert_and_test(model, input_np, verbose=False, change_ordering=change_ordering)
+    error = convert_and_test(model, input_np, verbose=False, should_transform_inputs=True)

test/utils.py

@@ -1,5 +1,4 @@
 import io
-import warnings
 
 import onnx
 import torch
@@ -9,6 +8,10 @@
 from onnx2kerastl import onnx_to_keras, check_torch_keras_error
 
 
+class LambdaLayerException(Exception):
+    pass
+
+
 def torch2keras(model: torch.nn.Module, input_variable, verbose=True, change_ordering=False):
     if isinstance(input_variable, (tuple, list)):
         input_variable = tuple(torch.FloatTensor(var) for var in input_variable)
@@ -36,7 +39,7 @@ def convert_and_test(model: torch.nn.Module,
     error = check_torch_keras_error(model, k_model, input_variable, change_ordering=change_ordering, epsilon=epsilon,
                                     should_transform_inputs=should_transform_inputs)
     if is_lambda_layers_exist(k_model):
-        warnings.warn("Found Lambda layers")
+        raise LambdaLayerException("Found Lambda layers")
     return error