pytorch转换onnx，再转换caffe 测试caffe, pytorch 模型结果是否一致

小灰灰 2023-01-03 04:47 148阅读 0赞

def pytorch_out(input):
        model = model_res() #model.eval
        # input = input.cuda()
        # model.cuda()
        torch.no_grad()
        t1 = time.time()
        output = model(input)
        print("一张图片torch 模型前向运行时间",time.time()-t1)
        # print output[0].flatten()[70:80]
        return output
    
    def caffe_output(input,input_layer_name="data",out_layer_name="fc1"):
        import sys
        # sys.path.insert(0, "/nfs-data/xingwg/deep_learning/NVCaffe/python")   #python3.6
        sys.path.insert(0, "/home/shiyy/nas/NVCaffe/python")  #python2.7
        import caffe
        caffe.set_device(0)
    
        deploy = "./transform_model/succed_res50.prototxt"
        weight = "./transform_model/succed_res50.caffemodel"
        caffe_model = caffe.Net(deploy, weight, caffe.TEST)
        # reshape network inputs
        blobs = {}
        blobs["0"] = input.data.numpy() #add dict
    
        t1=time.time()
        caffe_model.blobs[input_layer_name].reshape(*blobs["0"].shape)  #'input.1'  prototxt  name
        print("caffe 模型测试一张图片前向时间：",time.time()-t1)
    
        # do forward
        forward_kwargs = {input_layer_name: blobs['0'].astype(np.float32, copy=False)}  #input data name
        output_blobs = caffe_model.forward_all(**forward_kwargs)
    
        return output_blobs[out_layer_name]
    
    def pytorch_caffe_test():
        # 测试数据
        torch.manual_seed(66)
        dummy_input = torch.randn(1, 3, 112, 112, device='cpu')
    
        print("================>")
        caffe_out = caffe_output(dummy_input)
        print(caffe_out)
        print(caffe_out.shape)
    
        print("================>")
        torch_out_res = pytorch_out(dummy_input).detach().numpy()
        print(torch_out_res)
        print(torch_out_res.shape)
    
        print("===================================>")
        print("输出结果验证小数点后五位是否正确,都变成一维np")
        torch_out_res = torch_out_res.flatten()
        caffe_out = caffe_out.flatten()
    
        pytor = np.array(torch_out_res,dtype="float32") #need to float32
        caff=np.array(caffe_out,dtype="float32")  ##need to float32
        np.testing.assert_almost_equal(pytor,caff, decimal=5)
        print("恭喜你 ^^，caffe 和 pytorch结果一致， Exported model has been executed decimal=5 and the result looks good!")