[Relay][Keras] Permute, Softmax support (apache#3618)

Author: yongwww

python/tvm/relay/frontend/keras.py

@@ -115,6 +115,9 @@ def _convert_activation(inexpr, keras_layer, _):
 
 def _convert_advanced_activation(inexpr, keras_layer, etab):
     act_type = type(keras_layer).__name__
+
+    if act_type == 'Softmax':
+        return _op.nn.softmax(inexpr, axis=1)
     if act_type == 'ReLU':
         if keras_layer.max_value:
             return _op.clip(inexpr, a_min=0., a_max=float(keras_layer.max_value))
@@ -160,6 +163,8 @@ def _convert_merge(inexpr, keras_layer, _):
             'Operator {} is not supported in frontend Keras.'.format(merge_type))
     return ret
 
+def _convert_permute(inexpr, keras_layer, _):
+    return _op.transpose(inexpr, axes=(0,) + keras_layer.dims)
 
 def _convert_dense(inexpr, keras_layer, etab):
     weightList = keras_layer.get_weights()
@@ -574,6 +579,7 @@ def _default_skip(inexpr, keras_layer, _): # pylint: disable=unused-argument
 _convert_map = {
     'Dense'                    : _convert_dense,
     'Activation'               : _convert_activation,
+    'Softmax'                  : _convert_advanced_activation,
     'ReLU'                     : _convert_advanced_activation,
     'LeakyReLU'                : _convert_advanced_activation,
     'PReLU'                    : _convert_advanced_activation,
@@ -620,7 +626,7 @@ def _default_skip(inexpr, keras_layer, _): # pylint: disable=unused-argument
     'Average'                : _convert_merge,
     'Maximum'                : _convert_merge,
     # 'Dot'                    : _convert_merge,
-    # 'Permute'                : _convert_permute,
+    'Permute'                : _convert_permute,
     # 'Embedding'              : _convert_embedding,
     # 'RepeatVector'           : _convert_repeat_vector,
 
@@ -632,11 +638,15 @@ def _default_skip(inexpr, keras_layer, _): # pylint: disable=unused-argument
 
 
 def _check_unsupported_layers(model):
+    missing_ops = set()
     for layer in model.layers:
         op_name = type(layer).__name__
         if op_name not in _convert_map:
-            raise tvm.error.OpNotImplemented(
-                'Operator {} is not supported in frontend Keras.'.format(op_name))
+            missing_ops.add(op_name)
+
+    if missing_ops:
+        raise NotImplementedError( \
+            "The following operators are not implemented: {}".format(missing_ops))
 
 
 def keras_op_to_relay(inexpr, keras_layer, outname, etab):

tests/python/frontend/keras/test_forward.py

@@ -73,7 +73,7 @@ def to_channels_last(arr):
 
 
 def test_forward_merge():
-    data = keras.layers.Input(shape=(32,32,3))
+    data = keras.layers.Input(shape=(32, 32, 3))
     x = keras.layers.Conv2D(8, (3, 3), padding="same")(data)
     y = keras.layers.Conv2D(8, (3, 3), padding="same")(x)
     z = keras.layers.Conv2D(8, (3, 3), padding="same")(y)
@@ -93,7 +93,7 @@ def test_forward_merge():
 
 
 def test_forward_activations():
-    data = keras.layers.Input(shape=(32,32,3))
+    data = keras.layers.Input(shape=(32, 32, 3))
     act_funcs = [keras.layers.Activation('softmax'),
                  keras.layers.Activation('softplus'),
                  keras.layers.Activation('relu'),
@@ -103,6 +103,7 @@ def test_forward_activations():
                  keras.layers.Activation('tanh'),
                  keras.layers.Activation('linear'),
                  keras.layers.Activation('selu'),
+                 keras.layers.Softmax(),
                  keras.layers.ReLU(),
                  keras.layers.ReLU(max_value=6.),
                  keras.layers.LeakyReLU(alpha=0.3),
@@ -116,13 +117,18 @@ def test_forward_activations():
 
 
 def test_forward_dense():
-    data = keras.layers.Input(shape=(32,32,1))
+    data = keras.layers.Input(shape=(32, 32, 1))
     x = keras.layers.Flatten()(data)
     x = keras.layers.Dropout(0.5)(x)
     x = keras.layers.Dense(10, activation='relu', kernel_initializer='uniform')(x)
     keras_model = keras.models.Model(data, x)
     verify_keras_frontend(keras_model)
 
+def test_forward_permute():
+    data = keras.layers.Input(shape=(2, 3, 4))
+    x = keras.layers.Permute([2, 3, 1])(data)
+    keras_model = keras.models.Model(data, x)
+    verify_keras_frontend(keras_model, need_transpose=False)
 
 def test_forward_sequential():
     keras_model = keras.models.Sequential([
@@ -136,7 +142,7 @@ def test_forward_sequential():
 
 
 def test_forward_pool():
-    data = keras.layers.Input(shape=(32,32,1))
+    data = keras.layers.Input(shape=(32, 32, 1))
     # maxpool
     x = keras.layers.MaxPooling2D((3, 3), strides=(1, 1), padding='same')(data)
     keras_model = keras.models.Model(data, x)
@@ -148,36 +154,36 @@ def test_forward_pool():
 
 
 def test_forward_conv():
-    data = keras.layers.Input(shape=(32,32,3))
-    conv_funcs = [keras.layers.Conv2D(filters=10, kernel_size=(3,3),
-                                      strides=(2,2), padding='same'),
-                  keras.layers.Conv2D(filters=10, kernel_size=(3,3),
-                                      dilation_rate=(2,2), padding='same'),
-                  keras.layers.DepthwiseConv2D(kernel_size=(3,3), padding='same'),
-                  keras.layers.Conv2DTranspose(filters=10, kernel_size=(3,3), padding='valid'),
-                  keras.layers.SeparableConv2D(filters=10, kernel_size=(3,3), padding='same')]
+    data = keras.layers.Input(shape=(32, 32, 3))
+    conv_funcs = [keras.layers.Conv2D(filters=10, kernel_size=(3, 3),
+                                      strides=(2, 2), padding='same'),
+                  keras.layers.Conv2D(filters=10, kernel_size=(3, 3),
+                                      dilation_rate=(2, 2), padding='same'),
+                  keras.layers.DepthwiseConv2D(kernel_size=(3, 3), padding='same'),
+                  keras.layers.Conv2DTranspose(filters=10, kernel_size=(3, 3), padding='valid'),
+                  keras.layers.SeparableConv2D(filters=10, kernel_size=(3, 3), padding='same')]
     for conv_func in conv_funcs:
         x = conv_func(data)
         keras_model = keras.models.Model(data, x)
         verify_keras_frontend(keras_model)
 
 
 def test_forward_upsample(interpolation='nearest'):
-    data = keras.layers.Input(shape=(32,32,3))
-    x = keras.layers.UpSampling2D(size=(3,3), interpolation=interpolation)(data)
+    data = keras.layers.Input(shape=(32, 32, 3))
+    x = keras.layers.UpSampling2D(size=(3, 3), interpolation=interpolation)(data)
     keras_model = keras.models.Model(data, x)
     verify_keras_frontend(keras_model)
 
 
 def test_forward_reshape():
-    data = keras.layers.Input(shape=(32,32,3))
-    x = keras.layers.Reshape(target_shape=(32,32,3))(data)
+    data = keras.layers.Input(shape=(32, 32, 3))
+    x = keras.layers.Reshape(target_shape=(32, 32, 3))(data)
     keras_model = keras.models.Model(data, x)
     verify_keras_frontend(keras_model)
 
 
 def test_forward_crop():
-    data = keras.layers.Input(shape=(32,32,3))
+    data = keras.layers.Input(shape=(32, 32, 3))
     x = keras.layers.Cropping2D(cropping=((1, 1), (1, 1)))(data)
     x = keras.layers.Cropping2D(cropping=(1, 1))(x)
     x = keras.layers.Cropping2D(cropping=1)(x)
@@ -190,8 +196,8 @@ def test_forward_crop():
 
 
 def test_forward_multi_inputs():
-    data1 = keras.layers.Input(shape=(32,32,3))
-    data2 = keras.layers.Input(shape=(32,32,3))
+    data1 = keras.layers.Input(shape=(32, 32, 3))
+    data2 = keras.layers.Input(shape=(32, 32, 3))
     x = keras.layers.Conv2D(8, (3, 3), padding="same")(data1)
     y = keras.layers.Conv2D(8, (3, 3), padding="same")(data2)
     z = keras.layers.Average()([x, y])
@@ -201,7 +207,7 @@ def test_forward_multi_inputs():
 
 
 def test_forward_multi_outputs():
-    data = keras.layers.Input(shape=(32,32,3))
+    data = keras.layers.Input(shape=(32, 32, 3))
     x = keras.layers.Conv2D(8, (3, 3), padding="same")(data)
     x = keras.layers.GlobalAveragePooling2D()(x)
     y = keras.layers.Conv2D(8, (3, 3), padding="same")(data)
@@ -212,7 +218,7 @@ def test_forward_multi_outputs():
 
 def test_forward_reuse_layers():
     # reuse conv2d
-    data = keras.layers.Input(shape=(32,32,3))
+    data = keras.layers.Input(shape=(32, 32, 3))
     conv2d = keras.layers.Conv2D(8, (3, 3), padding="same")
     x = conv2d(data)
     y = conv2d(data)
@@ -221,7 +227,7 @@ def test_forward_reuse_layers():
     keras_model = keras.models.Model(data, z)
     verify_keras_frontend(keras_model)
     # reuse add
-    data = keras.layers.Input(shape=(32,32,3))
+    data = keras.layers.Input(shape=(32, 32, 3))
     x = keras.layers.Conv2D(8, (3, 3), padding="same")(data)
     add = keras.layers.Add()
     x = add([x, x])
@@ -232,7 +238,7 @@ def test_forward_reuse_layers():
 
 
 def test_forward_rnn():
-    data = keras.layers.Input(shape=(1,32))
+    data = keras.layers.Input(shape=(1, 32))
     rnn_funcs = [keras.layers.LSTM(units=16, return_state=False,
                     recurrent_activation='sigmoid', activation='tanh'),
                  keras.layers.SimpleRNN(units=16, return_state=False,
@@ -247,32 +253,33 @@ def test_forward_rnn():
 
 def test_forward_vgg16():
     keras_model = keras.applications.VGG16(include_top=True, weights='imagenet',
-        input_shape=(224,224,3), classes=1000)
+        input_shape=(224, 224, 3), classes=1000)
     verify_keras_frontend(keras_model)
 
 
 def test_forward_xception():
     keras_model = keras.applications.Xception(include_top=True, weights='imagenet',
-        input_shape=(299,299,3), classes=1000)
+        input_shape=(299, 299, 3), classes=1000)
     verify_keras_frontend(keras_model)
 
 
 def test_forward_resnet50():
     keras_model = keras.applications.ResNet50(include_top=True, weights='imagenet',
-        input_shape=(224,224,3), classes=1000)
+        input_shape=(224, 224, 3), classes=1000)
     verify_keras_frontend(keras_model)
 
 
 def test_forward_mobilenet():
     keras_model = keras.applications.MobileNet(include_top=True, weights='imagenet',
-        input_shape=(224,224,3), classes=1000)
+        input_shape=(224, 224, 3), classes=1000)
     verify_keras_frontend(keras_model)
 
 
 if __name__ == '__main__':
     test_forward_merge()
     test_forward_activations()
     test_forward_dense()
+    test_forward_permute()
     test_forward_sequential()
     test_forward_pool()
     test_forward_conv()