mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-05-15 20:50:42 +00:00
warn that initializers are in graph input provide a tool to move initializer out of graph input Motivation and Context ONNX model from IR_VERSION 4 only treats initializers that appear in graph input as non-constant. This may fail some of the graph optimizations, like const folding, operator fusion and etc. Warn the case and provide a tool.
37 lines
916 B
Python
37 lines
916 B
Python
import onnx
|
|
import sys
|
|
import argparse
|
|
|
|
|
|
def get_args():
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--input", required=True, help="input model")
|
|
parser.add_argument("--output", required=True, help="output model")
|
|
args = parser.parse_args()
|
|
return args
|
|
|
|
|
|
def remove_initializer_from_input():
|
|
args = get_args()
|
|
|
|
model = onnx.load(args.input)
|
|
if model.ir_version < 4:
|
|
print(
|
|
'Model with ir_version below 4 requires to include initilizer in graph input'
|
|
)
|
|
return
|
|
|
|
inputs = model.graph.input
|
|
name_to_input = {}
|
|
for input in inputs:
|
|
name_to_input[input.name] = input
|
|
|
|
for initializer in model.graph.initializer:
|
|
if initializer.name in name_to_input:
|
|
inputs.remove(name_to_input[initializer.name])
|
|
|
|
onnx.save(model, args.output)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
remove_initializer_from_input()
|