Deconvolution error

Gluon
1

I am trying to use a deconvolution layer or a Conv2DTranspose, but this error occurs.
Code:
self.up_sample = nn.Conv2DTranspose(c, kernel_size=3, strides=2, padding=1, groups=c)
Error:
src/operator/nn/deconvolution.cc:192: Check failed: dshape_nchw[1] % param_.num_group == 0U (255 vs. 0) : input num_filter must divide group size

I checked c right before the line and it was not 255 but 320. Also c divides c in any case. Here my understanding is that dshape_nchw[1] stands for the value of channel. Am I understanding the error in a wrong way? What should I do to avoid the error?

2

Hi, here c must divide the number of input channels, not the number of output channels. I believe if c devides the number of input channels (generally different than c), you should be fine.

Regards,
Foivos

1 Like