def onnx_graph_to_tensorflow_net(cls, graph_def, opset):
model_graph = tf.Graph()
with model_graph.as_default():
# initializer: TensorProtos representing the values to initialize
# a given tensor.
# initialized: A list of names of the initialized tensors.
if graph_def.initializer:
input_dict_items = cls.onnx_initializer_to_input_dict_items(
graph_def.initializer)
initialized = {init.name for init in graph_def.initializer}
else:
input_dict_items = []
initialized = set()
predict_net = TensorflowNet()
predict_net.name = graph_def.name
predict_net.graph = model_graph
predict_net.external_input.extend(
value_info.name for value_info in graph_def.input
if value_info.name not in initialized)
predict_net.external_output.extend(
value_info.name for value_info in graph_def.output)
# creating placeholders for currently unkown inputs
for value_info in graph_def.input:
if value_info.name in initialized:
continue
shape = list(d.dim_value if d.dim_value >= 0 else None
for d in value_info.type.tensor_type.shape.dim)
x = tf.placeholder(
TF_TYPE_ENUM[value_info.type.tensor_type.elem_type],
name=value_info.name,
shape=shape)
input_dict_items.append([value_info.name, x])
# tensor dict: this dictionary is a map from variable names
# to the latest produced TF tensors of the given name.
# This dictionary will get updated as we build the graph to
# record the names of newly produced tensors.
tensor_dict = dict(input_dict_items)
# Since tensor dict may be updated, we need to keep a copy
# of the original input dict where we track the earliest
# defined tensors so we can have access to the placeholders
# to feed in input tensors when we run the graph.
input_dict = dict(input_dict_items)
for node in graph_def.node:
node = OnnxNode(node)
output_ops = cls._onnx_node_to_tensorflow_op(node,
tensor_dict,
opset=opset)
curr_node_output_map = list(zip(node.outputs, output_ops))
tensor_dict = dict(
list(tensor_dict.items()) + curr_node_output_map)
predict_net.tensor_dict = tensor_dict
return predict_net
def onnx_graph_to_tensorflow_net(cls, graph_def):
# initializer: TensorProtos representing the values to initialize
# a given tensor.
# initialized: A list of names of the initialized tensors.
if graph_def.initializer:
input_dict_items = cls.onnx_initializer_to_input_dict_items(
graph_def.initializer)
initialized = {init.name for init in graph_def.initializer}
else:
input_dict_items = []
initialized = set()
predict_net = TensorflowNet()
predict_net.name = graph_def.name
predict_net.external_input.extend(value_info.name
for value_info in graph_def.input)
predict_net.external_output.extend(value_info.name
for value_info in graph_def.output)
# creating placeholders for currently unkown inputs
for value_info in graph_def.input:
if value_info.name in initialized:
continue
shape = list(d.dim_value
for d in value_info.type.tensor_type.shape.dim)
x = tf.placeholder(
cls.tensor_type_enum[value_info.type.tensor_type.elem_type],
name=value_info.name,
shape=shape)
input_dict_items.append([value_info.name, x])
# input dict: this dictionary is a map from variable names
# to the latest produced tensors of the given name.
# This dictionary will get updated as build the graph because
# some ops may produce a result tensor with the same name as
# the input tensor. The input dict tracks the latest produced
# tensors.
input_dict = dict(input_dict_items)
# Since input dict may be updated, we need to keep a copy
# of the original input dict where we track the earliest
# defined tensors so we can have access to the placeholders
# to feed in input tensors when we run the graph.
original_input_dict = dict(input_dict_items)
output_dict = dict()
for node in graph_def.node:
node = OnnxNode(node)
output_ops = cls._onnx_node_to_tensorflow_op(node, input_dict)
curr_node_output_map = list(zip(node.outputs, output_ops))
input_dict = dict(list(input_dict.items()) + curr_node_output_map)
output_dict = dict(
list(output_dict.items()) + curr_node_output_map)
predict_net.op.extend(output_ops)
predict_net.output_dict = output_dict
return original_input_dict, predict_net