[Opset 10] Updating ThresholdedRelu and onnxconverter-common package (#96)

- Updated onnxconverter-common versioning to 1.5.0
- ThresholdedRelu graduated from experimental op to full op in Opset 10 -- adding support for this new change and including test to ensure backwards compatibility

Author: vinitra

keras2onnx/ke2onnx/adv_activation.py

@@ -5,7 +5,7 @@
 ###############################################################################
 from ..proto import keras
 from distutils.version import StrictVersion
-from ..common.onnx_ops import apply_elu, apply_leaky_relu, apply_prelu
+from ..common.onnx_ops import apply_elu, apply_leaky_relu, apply_prelu, apply_thresholded_relu
 
 
 activations = keras.layers.advanced_activations
@@ -25,14 +25,15 @@ def convert_keras_advanced_activation(scope, operator, container):
         weights = op.get_weights()[0]
         apply_prelu(scope, operator.input_full_names[0], operator.output_full_names[0], container,
                     operator_name=operator.full_name, slope=weights)
+    elif isinstance(op, activations.ThresholdedReLU):
+        alpha = op.get_config()['theta']
+        apply_thresholded_relu(scope, operator.input_full_names[0], operator.output_full_names[0], container,
+                    operator_name=operator.full_name, alpha=[alpha])
     else:
         attrs = {'name': operator.full_name}
         ver_opset = 6
         input_tensor_names = [operator.input_full_names[0]]
-        if isinstance(op, activations.ThresholdedReLU):
-            op_type = 'ThresholdedRelu'
-            attrs['alpha'] = op.get_config()['theta']
-        elif StrictVersion(keras.__version__) >= StrictVersion('2.1.3') and \
+        if StrictVersion(keras.__version__) >= StrictVersion('2.1.3') and \
                 isinstance(op, activations.Softmax):
             op_type = 'Softmax'
             attrs['axis'] = op.get_config()['axis']
@@ -44,4 +45,4 @@ def convert_keras_advanced_activation(scope, operator, container):
         else:
             raise RuntimeError('Unsupported advanced layer found %s' % type(op))
 
-        container.add_node(op_type, input_tensor_names, operator.output_full_names, op_version=ver_opset, **attrs)
+        container.add_node(op_type, input_tensor_names, operator.output_full_names, op_version=ver_opset, **attrs)

requirements.txt

@@ -3,4 +3,4 @@ protobuf
 keras
 requests
 onnx
-onnxconverter-common>=1.4.0
+onnxconverter-common>=1.5.0

tests/test_layers.py

@@ -9,7 +9,7 @@
 import unittest
 import keras2onnx
 import numpy as np
-from keras2onnx.proto import keras, is_tf_keras
+from keras2onnx.proto import keras, is_tf_keras, get_opset_number_from_onnx
 from distutils.version import StrictVersion
 
 
@@ -382,7 +382,9 @@ def test_pooling_3d(self):
     def test_pooling_global(self):
         self._pooling_test_helper(keras.layers.GlobalAveragePooling2D, (4, 6, 2))
 
-    def activationlayer_helper(self, layer, data_for_advanced_layer=None):
+    def activationlayer_helper(self, layer, data_for_advanced_layer=None, op_version=None):
+        if op_version is None:
+            op_version = get_opset_number_from_onnx()
         if data_for_advanced_layer is None:
             data = self.asarray(-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5)
             layer = keras.layers.Activation(layer, input_shape=(data.size,))
@@ -391,7 +393,7 @@ def activationlayer_helper(self, layer, data_for_advanced_layer=None):
 
         model = keras.Sequential()
         model.add(layer)
-        onnx_model = keras2onnx.convert_keras(model, model.name)
+        onnx_model = keras2onnx.convert_keras(model, model.name, target_opset=op_version)
 
         expected = model.predict(data)
         self.assertTrue(self.run_onnx_runtime(onnx_model.graph.name, onnx_model, data, expected))
@@ -444,6 +446,8 @@ def test_LeakyRelu(self):
     def test_ThresholdedRelu(self):
         data = self.asarray(-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5)
         layer = keras.layers.advanced_activations.ThresholdedReLU(theta=1.0, input_shape=(data.size,))
+        self.activationlayer_helper(layer, data, op_version=8)
+        layer = keras.layers.advanced_activations.ThresholdedReLU(theta=1.0, input_shape=(data.size,))
         self.activationlayer_helper(layer, data)
 
     def test_ELU(self):