Error in converting mxnet to symbol model

Question

I use convolution instead of the shuffle operation in net

class ShuffleChannel(HybridBlock):

    def __init__(self, groups):
        super(ShuffleChannel, self).__init__()
        self.groups = groups

    def hybrid_forward(self, F, x):
        # shuffleOp
        # x.reshape((0, -4, self.groups, -1, -2)).swapaxes(1, 2).reshape((0, -3, -2))
        # return x
       
        # MineOp
        N, C, H, W = x.shape
        channels_per_group = C // self.groups
        conv_kernel = nd.zeros((C, C, 1, 1))
        for k in range(C):
            index = (k % self.groups) * channels_per_group + k // self.groups
            conv_kernel[k, index, 0, 0] = 1

        return nd.Convolution(x, conv_kernel, no_bias=True, kernel=(1,1), num_filter=C)

In training process, it works well and I want to convert the model to symbol format.But I got Errors:

......
  File "E:\AntiSpoofing\shuffleNetv2-mxnet\shufflenetv2.py", line 27, in hybrid_forward
    N, C, H, W = x.shape
AttributeError: 'Symbol' object has no attribute 'shape'

Could I specify the input 'x' to be in ndarray format or change the func code?

https://github.com/Rainweic/ShuffleNet-V2-MXNET/blob/43bf652a28a6812899c8f962a9dce6a02f80070f/shufflenetv2.py — Yuan Chu, Oct 14 '20 at 02:20

score 0 · Answer 1 · answered Oct 20 '20 at 22:33

Your code in hybrid_forward() is not "hybridizable", that is it can't be translated into symbolic graph. Symbol doesn't have shape property, but later you are constructing convolution filter. I'm unsure what you wish to achieve, but if you know the shapes apriori, you can initialize filter in init and use in hybrid_forward()

Error in converting mxnet to symbol model

1 Answers1