前言
构建onnx方式通常有两种：
1、通过代码转换成onnx结构，比如pytorch —> onnx
2、通过onnx 自定义结点，图，生成onnx结构
本文主要是简单学习和使用两种不同onnx结构，
下面以 Less 结点进行分析
方式
方法一：pytorch --> onnx
暂缓，主要研究方式二
方法二： onnx
import onnx
from onnx import TensorProto, helper, numpy_helper
import numpy as npdef run():print("run start....\n")less = helper.make_node("Less",name="Less_0",inputs=["input1", "input2"],outputs=["output1"],)input1_data = np.load("./tensor.npy") # 16, 397# input1_data = np.load("./data.npy") # 16, 398 test# print(f"input1_data shape:{input1_data.shape}\n")# input1_data = np.zeros((16,398))initializer = [ helper.make_tensor("input1", TensorProto.FLOAT, [16,397], input1_data)]cast_nodel = helper.make_node(op_type="Cast",inputs=["output1"],outputs=["output2"],name="test_cast",to=TensorProto.FLOAT,)value_info = helper.make_tensor_value_info("output2", TensorProto.BOOL, [16,397])graph = helper.make_graph(nodes=[less, cast_nodel],name="test_graph",inputs=[helper.make_tensor_value_info("input2", TensorProto.FLOAT, [16,1])],outputs=[helper.make_tensor_value_info("output2",TensorProto.FLOAT, [16,397])],initializer=initializer,value_info=[value_info],)op = onnx.OperatorSetIdProto()op.version = 11model = helper.make_model(graph, opset_imports=[op])model.ir_version = 8print("run done....\n")return modelif __name__ == "__main__":model = run()onnx.save(model, "./test_less_ori.onnx")
run
import onnx
import onnxruntime
import numpy as np# 检查onnx计算图
def check_onnx(mdoel):onnx.checker.check_model(model)# print(onnx.helper.printable_graph(model.graph))def run(model):print(f'run start....\n')session = onnxruntime.InferenceSession(model,providers=['CPUExecutionProvider'])input_name1 = session.get_inputs()[0].name input_data1= np.random.randn(16,1).astype(np.float32)print(f'input_data1 shape:{input_data1.shape}\n')output_name1 = session.get_outputs()[0].namepred_onx = session.run([output_name1], {input_name1: input_data1})[0]print(f'pred_onx shape:{pred_onx.shape} \n')print(f'run end....\n')if __name__ == '__main__':path = "./test_less_ori.onnx"model = onnx.load("./test_less_ori.onnx")check_onnx(model)run(path)
