def optimize_fp16_onnx_with_cast(input_onnx_path, optimized_onnx_path,
epsilon):
m = onnx.load(input_onnx_path)
onnx_model = OnnxModel(m)
nodes_to_remove = onnx_model.nodes()
nodes_to_add = [
onnx.helper.make_node("Cast", ["input"], ["fp32_input"],
"cast_input",
to=1),
onnx.helper.make_node("Cast", ["layer_norm.weight"],
["fp32_layer_norm.weight"],
"cast_weight",
to=1),
onnx.helper.make_node("Cast", ["layer_norm.bias"],
["fp32_layer_norm.bias"],
"cast_bias",
to=1),
onnx.helper.make_node(
"LayerNormalization",
["fp32_input", "fp32_layer_norm.weight", "fp32_layer_norm.bias"],
["fp32_output"],
"layer_norm",
epsilon=epsilon), # use fp32 epsilon
onnx.helper.make_node("Cast", ["fp32_output"], ["output"],
"cast_output",
to=10)
]
onnx_model.remove_nodes(nodes_to_remove)
onnx_model.add_nodes(nodes_to_add)
onnx_model.prune_graph()
onnx_model.save_model_to_file(optimized_onnx_path)
def optimize_fp16_onnx_with_cast(input_onnx_path, optimized_onnx_path,
epsilon):
m = onnx.load(input_onnx_path)
onnx_model = OnnxModel(m)
weight_name = get_weight(onnx_model)
bias_name = get_bias(onnx_model)
nodes_to_remove = [
n for n in onnx_model.nodes()
if n.output[0] != weight_name and n.output[0] != bias_name
]
nodes_to_add = [
onnx.helper.make_node("Cast", ["input"], ["fp32_input"],
"cast_input",
to=1),
onnx.helper.make_node("Cast", [weight_name],
["fp32_layer_norm.weight"],
"cast_weight",
to=1),
onnx.helper.make_node("Cast", [bias_name], ["fp32_layer_norm.bias"],
"cast_bias",
to=1),
onnx.helper.make_node(
"LayerNormalization",
["fp32_input", "fp32_layer_norm.weight", "fp32_layer_norm.bias"],
["fp32_output"],
"layer_norm",
epsilon=epsilon,
), # use fp32 epsilon
onnx.helper.make_node("Cast", ["fp32_output"], ["output"],
"cast_output",
to=10),
]
onnx_model.remove_nodes(nodes_to_remove)
onnx_model.add_nodes(nodes_to_add)
onnx_model.prune_graph()
onnx_model.save_model_to_file(optimized_onnx_path)