pytorch转为onnx格式：

def Torch2Onnx(model,input_size,output_name,istrained=True):
    '''
    :param: model
    :param: input_size .e.t. (244,244)
    :param: output_name .e.t. "test_output"
    :param: if convert a trained model or not. default: True
    '''
    x = Variable(torch.randn(1,3,input_size[0],input_size[1])).cuda()
    if istrained:
        torch_out = torch.onnx.export(model,x,output_name,verbose=True)
    else:
        torch_out = torch.onnx.export(model,x,output_name,export_params=False,verbose=True) # Only export a untrained model.

使用举例：

model = model()
model.load_state_dict(torch.load(weight_path))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
input_size = (384,288)
Torch2Onnx(model,input_size,"test.onnx")

获取model中的params：

请注意：不同的方法默认model在cpu还是在cuda上是不一样的，如果出现类似RuntimeError: Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same的报错，请检查weight是否应该在cuda上。

方法一：使用torchsummary

1. 使用pip安装torchsummary：
   pip install torchsummary

2. 代码片段：

   from torchsummary import summary
   model = model()
   model.load_state_dict(torch.load(weight_path))
   device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
   model = model.to(device)
   summary(model,(3,384,288))
   

方法二：使用torchstat

1. 使用pip安装torchstat：
   pip install torchstat

2. 代码片段（和summary差不多）

   from torchstat import stat
    model = model()
    model.load_state_dict(torch.load(weight_path))
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    stat(model,(3,384,288))
   

方法三：使用thop（不太推荐）

1. 使用pip安装thop：
   pip install thop

2. 代码片段：

   from thop import profile,clever_format
   model = model()
    model.load_state_dict(torch.load(weight_path))
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    flops, params = profile(model,inputs=())
    flops,params = clever_format(flops,params,"%.3f")