Generate graph after optimizing onnx model (#34)

* Generate graph after optimizing onnx

onnxconverter_common/__init__.py

@@ -9,7 +9,7 @@
 This framework performs optimization for ONNX models and
 includes common utilities for ONNX converters.
 """
-__version__ = "1.6.0"
+__version__ = "1.6.1"
 __author__ = "Microsoft"
 __producer__ = "OnnxMLTools"
 __producer_version__ = __version__

onnxconverter_common/optimizer.py

@@ -789,6 +789,63 @@ def optimize_onnx(onnx_nodes, nchw_inputs=None, inputs=None, outputs=None):
     return _build_onnx_model(node_list)
 
 
+def _remove_unused_initializers(nodes, initializers):
+    adjusted_initializers = []
+    nodes_input_set = set()
+    for n_ in nodes:
+        for input_name_ in n_.input:
+            nodes_input_set.add(input_name_)
+
+    for initializers_ in initializers:
+        if initializers_.name in nodes_input_set:
+            adjusted_initializers.append(initializers_)
+
+    return adjusted_initializers
+
+
+def optimize_onnx_graph(onnx_nodes, nchw_inputs=None, inputs=None, outputs=None, initializers=None,
+                        model_value_info=None, model_name='', target_opset=None):
+    """
+    Optimize onnx model by several approaches.
+    :param onnx_nodes: the onnx node list in onnx model.
+    :param nchw_inputs: the name list of the inputs needed to be transposed as NCHW
+    :param inputs: the model input
+    :param outputs: the model output
+    :param initializers: the model initializers
+    :param model_value_info: the model value_info
+    :return: the optimized onnx graph
+    """
+    if target_opset < 9:
+        raise Exception("target_opset = {}, Use optimize_onnx_graph for opset >= 9".format(target_opset))
+
+    # When calling ModelComponentContainer's add_initializer(...), nothing is added into the input list.
+    # However, In ONNX, for target opset < 9, initializers should also be model's (GraphProto) inputs.
+    # Thus, we create ValueInfoProto objects from initializers (type: TensorProto) directly and then add them into model's input list.
+    extra_inputs = []  # ValueInfoProto list of the initializers
+    for tensor in initializers:
+        # Sometimes (especially when creating optional input values such as RNN's initial hidden state), an initializer
+        # is also one of the original model's input, so it has been added into the container's input list. If this is
+        # the case, we need to skip one iteration to avoid duplicated inputs.
+        if tensor.name in [value_info.name for value_info in inputs]:
+            continue
+
+        # Initializers are always tensors so we can just call make_tensor_value_info(...)
+        value_info = helper.make_tensor_value_info(tensor.name, tensor.data_type, tensor.dims)
+        extra_inputs.append(value_info)
+
+    nodes = optimize_onnx(onnx_nodes, nchw_inputs=nchw_inputs, inputs=inputs + extra_inputs, outputs=outputs)
+
+    # Create a graph from its main components
+    adjusted_initializers = _remove_unused_initializers(nodes, initializers)
+    graph = helper.make_graph(nodes, model_name, inputs,
+                              outputs, adjusted_initializers)
+
+    # Add extra information related to the graph
+    graph.value_info.extend(model_value_info)
+
+    return graph
+
+
 def optimize_onnx_model(origin_model, nchw_inputs=None):
     """
     the origin model will be updated after the optimization.