Upgrade to opset 14 (#723)

Author: jiafatom

.azure-pipelines/linux-CI-keras-applications-nightly.yml

@@ -38,7 +38,7 @@ jobs:
 
       Python37:
         python.version: '3.7.3'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         INSTALL_KERAS: pip install keras==2.3.1
         UNINSTALL_KERAS:
         INSTALL_TENSORFLOW: pip install tensorflow==1.15.0
@@ -60,7 +60,7 @@ jobs:
 
       Python38-tf2:
         python.version: '3.8'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         INSTALL_KERAS:
         UNINSTALL_KERAS: pip uninstall keras -y
         INSTALL_TENSORFLOW: pip install tensorflow==2.2.0

.azure-pipelines/linux-conda-CI-tf-keras.yml

@@ -36,13 +36,13 @@ jobs:
         python.version: '3.8'
         ONNX_PATH: onnx==1.7.0
         TENSORFLOW_PATH: tensorflow-cpu==2.2.0
-        INSTALL_ORT:
+        INSTALL_ORT: pip install onnxruntime==1.6.0
 
       Python38-tf2.3:
         python.version: '3.8'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         TENSORFLOW_PATH: tensorflow-cpu==2.3.0
-        INSTALL_ORT: pip install onnxruntime==1.7.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
 
     maxParallel: 3
 

.azure-pipelines/linux-conda-CI.yml

@@ -29,10 +29,10 @@ jobs:
 
       Python37-tf1150:
         python.version: '3.7'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         KERAS: keras==2.4.3
         TENSORFLOW_PATH: tensorflow==1.15.0
-        INSTALL_ORT: pip install onnxruntime==1.6.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
 
       # UT for keras 2.3 need tensorflow <= 2.0.0 
       Python37-tf200:

.azure-pipelines/win32-CI-keras-applications-nightly.yml

@@ -39,7 +39,7 @@ jobs:
 
       Python37:
         python.version: '3.7'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         INSTALL_KERAS: pip install keras==2.3.1
         UNINSTALL_KERAS:
         INSTALL_TENSORFLOW: pip install tensorflow==1.14.0
@@ -61,11 +61,11 @@ jobs:
 
       Python38-tf2:
         python.version: '3.8'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         INSTALL_KERAS:
         UNINSTALL_KERAS: pip uninstall keras -y
         INSTALL_TENSORFLOW: pip install tensorflow==2.2.0
-        INSTALL_ORT: pip install onnxruntime==1.7.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
         INSTALL_KERAS_RESNET: pip install keras-resnet
         INSTALL_TRANSFORMERS: pip install transformers==3.4.0
         NIGHTLY_BUILD_TEST: python run_all_v2.py

.azure-pipelines/win32-conda-CI-tf-keras.yml

@@ -36,13 +36,13 @@ jobs:
         python.version: '3.8'
         ONNX_PATH: onnx==1.7.0
         TENSORFLOW_PATH: tensorflow-cpu==2.2.0
-        INSTALL_ORT:
+        INSTALL_ORT: pip install onnxruntime==1.6.0
 
       Python38-tf2.3:
         python.version: '3.8'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         TENSORFLOW_PATH: tensorflow-cpu==2.3.0
-        INSTALL_ORT: pip install onnxruntime==1.7.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
 
     maxParallel: 3
 

.azure-pipelines/win32-conda-CI.yml

@@ -38,10 +38,10 @@ jobs:
 
       Python38-tf220:
         python.version: '3.8'
-        ONNX_PATH: onnx==1.8.0
+        ONNX_PATH: onnx==1.9.0
         KERAS: keras==2.4.3
         TENSORFLOW_PATH: tensorflow==2.2.0
-        INSTALL_ORT:
+        INSTALL_ORT: pip install onnxruntime==1.8.0
 
     maxParallel: 3
 

keras2onnx/__init__.py

@@ -4,7 +4,7 @@
 keras2onnx
 This package converts keras models into ONNX for use with any inference engine supporting ONNX
 """
-__version__ = "1.8.0"
+__version__ = "1.9.0"
 __author__ = "Microsoft Corporation"
 __producer__ = "keras2onnx"
 

keras2onnx/_builtin.py

@@ -796,7 +796,8 @@ def _convert_tf_fused_batch_norm_core(scope, operator, container):
     input_dim = len(_cal_tensor_shape(node.inputs[0]))
     epsilon = node.get_attr('epsilon')
     attrs = {'epsilon': epsilon, 'momentum': 0.9, 'spatial': 1}
-    outputs_num = min(5, len(node.outputs))
+    # ORT assumes opitonal outputs indicate training mode. So we should use one output for inference.
+    outputs_num = 1
 
     if _is_nhwc(node):
         input_perm = [0, input_dim - 1] + list(range(1, input_dim - 1))

keras2onnx/ke2onnx/batch_norm.py

@@ -65,7 +65,8 @@ def convert_keras_batch_normalization(scope, operator, container):
 
     if skip_transpose:
         # If no transpose is required, we can simply use the output of ONNX BatchNorm as the final outcome
-        apply_batch_norm(scope, input_tensor_names, operator.output_full_names[0:5], container,
+        # ORT assumes opitonal outputs indicate training mode. So we should use one output for inference.
+        apply_batch_norm(scope, input_tensor_names, operator.output_full_names[0], container,
                          operator_name=operator.full_name, epsilon=epsilon, is_test=is_test,
                          momentum=momentum, spatial=spatial)
     else:

requirements.txt

@@ -2,5 +2,5 @@ numpy
 protobuf
 requests
 onnx
-onnxconverter-common>=1.8.0
+onnxconverter-common>=1.9.0
 fire