SR3 Natural Image (ImageNet) Super-Resolution Reproduction

[abhi1kumar]

Hi,

We have successfully reproduced the PSNR and SSIM numbers for the FFHQ-CelebaHQ super-resolution experiments (32x32 -> 128x128) using batch size of 256 for 1M iterations. However, we are running into some challenges reproducing the natural image super-resolution (64x64 -> 256x256) numbers from the paper. We made these adjustments following the SR3 paper for natural image super-resolution experiment:

• Data Preparation: We filtered ImageNet images to ensure the minimum height/width was at least 256 pixels.
• Architecture: We added more depth multipliers as outlined in Table A.1 of the paper.
• Dropout: We changed the dropout to 0.0, following the training details section in the paper.

We trained the model on 0.1M ImageNet training images with batch size of 112, The reproduced PSNR is around 14-16, while the SSIM is around 0.41, which are much lower than the reported numbers of 26.4 PSNR in the paper. (With 0.1M images, I was expecting atleast 20 PSNR)

	#Train Im	Batch	PSNR
Reported	1.2M	256	26.4
Reproduced	0.1M	112	14.0

It would be great if you could suggest changes needed to reproduce ImageNet experiment numbers

[zmy-123-zmz]

我也是复现出来指标很低，训练到30万次psnr才只有11点多，能询问一下你做的4倍超分指标值是使用了这篇文献代码得到的结果嘛？