Questions about training hyperparameters

[zzukong]

Hello author, thank you very much for your excellent work. Recently while reproducing your code, I realized that no amount of training could achieve the excellent performance with the weights you provided. I did the training with NVIDIA GeForce 3090 GPUs and did not change any hyperparameters other than batchsize. In total, I tried setting the batchsize to 16 (using two graphics cards), 6 (as provided in the paper) and 2 (the source code default). Found that the smaller the batchsize the better the performance seems to be, but only up to 23psnr. This is a big difference from the weights you provided, could you please tell me some specifics about the training?

[jrcyyzb]

我用raw图训练，最后调整的结果有马赛克的情况

[Li-Chongyi]

建议使用RGB图像，原生图像效果没有试过，可能不适用。

jrcyyzb @.***> 于2023年8月29日周二 13:50写道：

我用raw图训练，最后调整的结果有马赛克的情况

— Reply to this email directly, view it on GitHub https://github.com/Li-Chongyi/UHDFour_code/issues/2#issuecomment-1696811771, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIEVQEK2GXGTIJE2KYWBN2TXXV7LTANCNFSM6AAAAAA3JIV7S4 . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[Li-Chongyi]

I just used the settings provided in the paper to produce the results. We used V100 to train the model.

zzukong @.***> 于2023年8月9日周三 09:31写道：

Hello author, thank you very much for your excellent work. Recently while reproducing your code, I realized that no amount of training could achieve the excellent performance with the weights you provided. I did the training with NVIDIA GeForce 3090 GPUs and did not change any hyperparameters other than batchsize. In total, I tried setting the batchsize to 16 (using two graphics cards), 6 (as provided in the paper) and 2 (the source code default). Found that the smaller the batchsize the better the performance seems to be, but only up to 23psnr. This is a big difference from the weights you provided, could you please tell me some specifics about the training?

— Reply to this email directly, view it on GitHub https://github.com/Li-Chongyi/UHDFour_code/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIEVQEK5KN46XRA4WJUXMCDXULR5PANCNFSM6AAAAAA3JIV7S4 . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[jrcyyzb]

Hello author, thank you very much for your excellent work. Recently while reproducing your code, I realized that no amount of training could achieve the excellent performance with the weights you provided. I did the training with NVIDIA GeForce 3090 GPUs and did not change any hyperparameters other than batchsize. In total, I tried setting the batchsize to 16 (using two graphics cards), 6 (as provided in the paper) and 2 (the source code default). Found that the smaller the batchsize the better the performance seems to be, but only up to 23psnr. This is a big difference from the weights you provided, could you please tell me some specifics about the training?

你是加载的作者的预训练权重吗，你一共训练了多少轮，我只能达到20psnr左右

[jrcyyzb]

建议使用RGB图像，原生图像效果没有试过，可能不适用。 jrcyyzb @.> 于2023年8月29日周二 13:50写道： … 我用raw图训练，最后调整的结果有马赛克的情况 — Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIEVQEK2GXGTIJE2KYWBN2TXXV7LTANCNFSM6AAAAAA3JIV7S4 . You are receiving this because you are subscribed to this thread.Message ID: @.>

论文里是全部的训练细节了吗，我的测试集结果还是有马赛克，请问你训练的时候出现过这种问题吗

[jrcyyzb]

Hello author, thank you very much for your excellent work. Recently while reproducing your code, I realized that no amount of training could achieve the excellent performance with the weights you provided. I did the training with NVIDIA GeForce 3090 GPUs and did not change any hyperparameters other than batchsize. In total, I tried setting the batchsize to 16 (using two graphics cards), 6 (as provided in the paper) and 2 (the source code default). Found that the smaller the batchsize the better the performance seems to be, but only up to 23psnr. This is a big difference from the weights you provided, could you please tell me some specifics about the training?

你是加载的作者的预训练权重吗，你一共训练了多少轮，我只能达到20psnr左右 而且奇怪的是我用作者提供的权重在测试集上只能得到24.6368psnr

[Li-Chongyi]

代码里面提供了完整的 训练和测试代码 而且是根据不同的数据集有不同的设置 直接可以训练和测试的 PSNR 可以查看一下 是否和代码提供的算法一直

jrcyyzb @.***> 于2023年9月26日周二 09:28写道：

Hello author, thank you very much for your excellent work. Recently while reproducing your code, I realized that no amount of training could achieve the excellent performance with the weights you provided. I did the training with NVIDIA GeForce 3090 GPUs and did not change any hyperparameters other than batchsize. In total, I tried setting the batchsize to 16 (using two graphics cards), 6 (as provided in the paper) and 2 (the source code default). Found that the smaller the batchsize the better the performance seems to be, but only up to 23psnr. This is a big difference from the weights you provided, could you please tell me some specifics about the training?

你是加载的作者的预训练权重吗，你一共训练了多少轮，我只能达到20psnr左右 而且奇怪的是我用作者提供的权重在测试集上只能得到24.6368psnr

— Reply to this email directly, view it on GitHub https://github.com/Li-Chongyi/UHDFour_code/issues/2#issuecomment-1734691917, or unsubscribe https://github.com/notifications/unsubscribe-auth/AIEVQEMIB3UBTNVQRXKIY73X4IVUBANCNFSM6AAAAAA3JIV7S4 . You are receiving this because you commented.Message ID: @.***>