Edit inputs from an ONNX model

[IzanCatalan]

Hi all!, Does anyone know how to edit inputs from an onnx model in python?

I would like to change for exameple one element of an input tensor (like weights in a convolution layer) in one node of the graph. I tried to do it with " model = onnx.load ()" and after with "inputs = model.graph.initializers" but when I get de inputs, I can read them but not write them, they're read only.

is there any way to write them, using another tool for instance? Thanks.

[tianleiwu]

You can write them directly (I mean change data member like data_type, raw_data etc. of initializer).

For example, we have a tool to convert model from float32 to float16. That need change initializers of float32. Related code:

onnxruntime/onnxruntime/python/tools/transformers/float16.py

Lines 68 to 83 in bdd6b00

	if tensor.data_type == onnx_proto.TensorProto.FLOAT:
	tensor.data_type = onnx_proto.TensorProto.FLOAT16
	# convert float_data (float type) to float16 and write to int32_data
	if tensor.float_data:
	float16_data = convert_np_to_float16(np.array(tensor.float_data), min_positive_val, max_finite_val)
	int_list = _npfloat16_to_int(float16_data)
	tensor.int32_data[:] = int_list
	tensor.float_data[:] = []
	# convert raw_data (bytes type)
	if tensor.raw_data:
	# convert n.raw_data to float
	float32_list = np.frombuffer(tensor.raw_data, dtype="float32")
	# convert float to float16
	float16_list = convert_np_to_float16(float32_list, min_positive_val, max_finite_val)
	# convert float16 to bytes and write back to raw_data
	tensor.raw_data = float16_list.tobytes()