search

caiyuanhao1998 / MST

A toolbox for spectral compressive imaging reconstruction including MST (CVPR 2022), CST (ECCV 2022), DAUHST (NeurIPS 2022), BiSCI (NeurIPS 2023), HDNet (CVPR 2022), MST++ (CVPRW 2022), etc.
MIT License
571 stars 71 forks source link

could you provide the training logs about DAUHST model? #19

Closed Junyu99 closed 1 year ago

Junyu99 commented 1 year ago

I am very interesed in your DAUHST work and reproducing your expermental work. Could you prvoid the training logs of DAUHST, it would be helpful to have training logs to inspect the correctness of the training process.

caiyuanhao1998 commented 1 year ago

What problem have you encountered?

Junyu99 commented 1 year ago

We have trained DAUHST-2stg, 3stg, 5stg model by using your provided source code and commands. And we got the results as follows: 2stg: Epoch 296: testing psnr = 36.35, ssim = 0.957, time: 0.05 (PSNR,SSIM: 36.34/0.952 in paper, +0.01/+0.005) 3stg: Epoch 277: testing psnr = 36.95, ssim = 0.963, time: 0.06 (PSNR,SSIM: 37.21/0.959 in paper, -0.26/+0.004) 5stg: Epoch 242: testing psnr = 37.62, ssim = 0.968, time: 0.22 (PSNR,SSIM: 37.75/0.962 in paper, -0.13/+0.006)

It is confusing to me that PSNR results of DAUHSR-3stg and 5stg model we have trained are worse than the results in your paper, but the SSIM results of DAUHST-3stg and 5stg model are better than the results in your paper.

Does the random seed affect the results so much?

caiyuanhao1998 commented 1 year ago

This is the relative training error within a reasonable range. It is OK. I suggest you to train several times if you are not satisfied.

Junyu99 commented 1 year ago

Alright, thanks for your reply.

caiyuanhao1998 commented 1 year ago

You are welcome