Training CNNs using real deconvolution techniques

[cxy]

The current way of training CNNs is likely to be inefficient and indirect. Can pytorch developers give a closer look to this issue?

Issue: Training CNNs with deconvolution