Shifted function from : FluxML/Flux.jl#735 , as discussed with @dhairyagandhi96 .

I have created a new file specifically for these sort of operations on convolutions as adding this to conv.jl din't seem to fit in right, as it is filled with optimization and convolution utility functions.

All such operations in future can be added to conv_op.jl. Is it fine?

Codecov Report

Merging #112 into master will decrease coverage by 3.29%.
The diff coverage is 0%.

@@            Coverage Diff            @@
##           master     #112     +/-   ##
=========================================
- Coverage   96.66%   93.36%   -3.3%     
=========================================
  Files          17       18      +1     
  Lines         539      558     +19     
=========================================
  Hits          521      521             
- Misses         18       37     +19

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 40cee4b...766f524. Read the comment docs.

I've also implemented a pixel shuffle, as described here: https://arxiv.org/abs/1609.05158

My implementation is different from yours (just a reshape and permutedims, no need to partition) and, on running the code, doesn't shuffle pixels precisely the same way. Do you have a reference describing in detail the operation implemented here?

I've followed the following for refernce :
https://github.com/brade31919/SRGAN-tensorflow/blob/7f7ec0f929feb64a111221dc992f740cd48e11eb/lib/ops.py#L85-L112

The partition is for splitting the channels so as to cyclically permute them.

@dhairyagandhi96 @staticfloat What else needs to be done more in this?

Is this feature available in Flux yet? Considering porting code from pytorch that uses this. Thanks!

superseded by #260