2021-10-09

出现了一个问题，下图是手动搭建的卷积，标准化，relu6的运算下载结果

代码如下

         out1 = F.conv2d(x, filter101, padding=(1,1),groups=1,stride= (2,2))


        # print(out.shape)得到[1,16,112,112]

        # 标准化 clip
        out = nn.BatchNorm2d(16, affine=True, eps = 0.0001, momentum= 0.1,                     track_running_stats=True)

        dataw = torch.as_tensor(data_w_101)

        out.weight = torch.nn.Parameter(dataw)
        datab = torch.as_tensor(data_b_101)
        
        out.bias = torch.nn.Parameter(datab)
        datam = torch.as_tensor((mean101))
        
        out.running_mean = datam
        datav = torch.as_tensor(var101)
        
        out.running_var = datav
        out1 = out(out1)
        
        print(out1.shape)
        print(out1)

        out1 = clip(out1)

下图是网路中第一层，也就是上图我自己搭建的过程的输出

 会发现有一些偏差，每一个值都有一些偏差

我自己单独搭建的卷积，标准化，relu6的过程所采取的参数是取自pytorch网络模型导出为onnx后得到的