Merge pull request #1547 from xadupre/ortext

Replace onnxruntime_customops by onnxruntime_extensions

Author: xadupre

README.md

@@ -293,6 +293,9 @@ model_proto, external_tensor_storage = tf2onnx.convert.from_keras(model,
         model: the tf.keras model we want to convert
         input_signature: a tf.TensorSpec or a numpy array defining the shape/dtype of the input
         opset: the opset to be used for the ONNX model, default is the latest
+        custom_ops: if a model contains ops not recognized by onnx runtime,
+            you can tag these ops with a custom op domain so that the
+            runtime can still open the model. Type is a dictionary `{op name: domain}`.
         target: list of workarounds applied to help certain platforms
         custom_op_handlers: dictionary of custom ops handlers
         custom_rewriter: list of custom graph rewriters
@@ -322,6 +325,9 @@ model_proto, external_tensor_storage = tf2onnx.convert.from_function(function,
         function: the tf.function we want to convert
         input_signature: a tf.TensorSpec or a numpy array defining the shape/dtype of the input
         opset: the opset to be used for the ONNX model, default is the latest
+        custom_ops: if a model contains ops not recognized by onnx runtime,
+            you can tag these ops with a custom op domain so that the
+            runtime can still open the model. Type is a dictionary `{op name: domain}`.
         target: list of workarounds applied to help certain platforms
         custom_op_handlers: dictionary of custom ops handlers
         custom_rewriter: list of custom graph rewriters

ci_build/azure_pipelines/templates/setup.yml

@@ -25,12 +25,20 @@ steps:
       pip install --index-url https://test.pypi.org/simple/ ort-nightly
     fi
 
-    if [[ $CI_PLATFORM == "windows" ]] ;
+    if [[ $CI_TF_VERSION == 2.* ]] ;
     then
-      pip install -i https://test.pypi.org/simple/ onnxruntime-customops==0.0.1
-      if [[ $CI_TF_VERSION == 2.* ]] ;
+      pip install onnxruntime-extensions
+      if [[ $CI_TF_VERSION == 2.3* ]] ;
       then
-        pip install tensorflow-text
+        pip install tensorflow-text==${CI_TF_VERSION}
+      fi
+      if [[ $CI_TF_VERSION == 2.4* ]] ;
+      then
+        pip install tensorflow-text==${CI_TF_VERSION}
+      fi
+      if [[ $CI_TF_VERSION == 2.5* ]] ;
+      then
+        pip install tensorflow-text>=2.5
       fi
     fi
 

tests/backend_test_base.py

@@ -80,7 +80,7 @@ def run_onnxruntime(self, model_path, inputs, output_names, use_custom_ops=False
                 providers = ['CUDAExecutionProvider']
         opt = rt.SessionOptions()
         if use_custom_ops:
-            from onnxruntime_customops import get_library_path
+            from onnxruntime_extensions import get_library_path
             opt.register_custom_ops_library(get_library_path())
         # in case of issues with the runtime, one can enable more logging
         # opt.log_severity_level = 0

tests/common.py

@@ -242,7 +242,7 @@ def requires_custom_ops(message=""):
     """ Skip until custom ops framework is on PyPI. """
     reason = _append_message("test needs custom ops framework", message)
     try:
-        import onnxruntime_customops  #pylint: disable=import-outside-toplevel,unused-import
+        import onnxruntime_extensions  #pylint: disable=import-outside-toplevel,unused-import
         can_import = True
     except ModuleNotFoundError:
         can_import = False

tf2onnx/convert.py

@@ -332,6 +332,9 @@ def from_keras(model, input_signature=None, opset=None, custom_ops=None, custom_
         model: the tf.keras model we want to convert
         input_signature: a tf.TensorSpec or a numpy array defining the shape/dtype of the input
         opset: the opset to be used for the ONNX model, default is the latest
+        custom_ops: if a model contains ops not recognized by onnx runtime,
+            you can tag these ops with a custom op domain so that the
+            runtime can still open the model. Type is a dictionary `{op name: domain}`.
         target: list of workarounds applied to help certain platforms
         custom_op_handlers: dictionary of custom ops handlers
         custom_rewriter: list of custom graph rewriters
@@ -417,6 +420,9 @@ def from_function(function, input_signature=None, opset=None, custom_ops=None, c
         function: the tf.function we want to convert
         input_signature: a tf.TensorSpec or a numpy array defining the shape/dtype of the input
         opset: the opset to be used for the ONNX model, default is the latest
+        custom_ops: if a model contains ops not recognized by onnx runtime,
+            you can tag these ops with a custom op domain so that the
+            runtime can still open the model. Type is a dictionary `{op name: domain}`.
         target: list of workarounds applied to help certain platforms
         custom_op_handlers: dictionary of custom ops handlers
         custom_rewriter: list of custom graph rewriters
@@ -478,6 +484,9 @@ def from_graph_def(graph_def, name=None, input_names=None, output_names=None, op
         output_names: list of output names
         name: A name for the graph
         opset: the opset to be used for the ONNX model, default is the latest
+        custom_ops: if a model contains ops not recognized by onnx runtime,
+            you can tag these ops with a custom op domain so that the
+            runtime can still open the model. Type is a dictionary `{op name: domain}`.
         target: list of workarounds applied to help certain platforms
         custom_op_handlers: dictionary of custom ops handlers
         custom_rewriter: list of custom graph rewriters