4.4 获取onnx每个节点的输出结果

nsq_ai2023-11-13 11:59

前言

获取onnx每个节点的结果，进行输出显示、保存

Code

cpp 复制代码

import os
import onnx
import onnx.helper as helper
import onnxruntime
from collections import OrderedDict
import  numpy as np

def get_onnx_node_out(onnx_file, save_onnx):
    model = onnx.load(onnx_file)
    out_names=[]
    for i, node in enumerate(model.graph.node):
        out_names.append(node.output[0])
    for out_name in out_names:
        intermediate_layer_value_info = helper.ValueInfoProto()
        intermediate_layer_value_info.name = out_name
        model.graph.output.append(intermediate_layer_value_info)
    onnx.save(model, save_onnx)

def onnxruntime_infer(onnx_path, input_data, output_name="output"):
 
    session = onnxruntime.InferenceSession(onnx_path, providers=['CPUExecutionProvider'])
    input_name = session.get_inputs()[0].name
    outputs = [x.name for x in session.get_outputs()]
    print("onnx input_name:", input_name)
    print("onnx outputs:", outputs)
    ort_outs = session.run(outputs, {input_name: input_data})
    ort_outs = OrderedDict(zip(outputs, ort_outs))
 
    # For debug
    for key in ort_outs:
        val = ort_outs[key]
        file = "./onnx_output/"+ key.split("/")[-1] +".npy"
        np.save(file, val, allow_pickle=True, fix_imports=True)
     
if __name__ == '__main__':

    base_path = "./"
    onnx_file = os.path.join(base_path,"example4.onnx")
    save_onnx = os.path.join(base_path,"example4_out.onnx")
    get_onnx_node_out(onnx_file, save_onnx)

    path = "./10.npy"  # 
    input_data = np.load(path)
    print(f"input_data shape:{input_data.shape}")

    onnxruntime_infer(save_onnx, input_data)

总结

相关代码简单运用