Can this method be used for networks whose outputs are images? I followed the example to implement pruning for an image restoration network, but after pruning the PSNR dropped sharply, and fine-tuning caused the PSNR to decrease even further. It seems that the gradient computation has become incorrect.
Can this method be used for networks whose outputs are images? I followed the example to implement pruning for an image restoration network, but after pruning the PSNR dropped sharply, and fine-tuning caused the PSNR to decrease even further. It seems that the gradient computation has become incorrect.