关于Flops

[WZX2022]

作者您好，我想请问一下论文中DRCT和HAT模型的Flops计算是基于怎样的输入尺寸呢

[xduzhangjiayu]

作者您好，我想请问一下论文中DRCT和HAT模型的Flops计算是基于怎样的输入尺寸呢

我发现同样的输入HAT要比DRCT快很多，不知道是不是我配置的问题？而且我发现DRCT保存下来的checkpoint特别大，很奇怪

[ming053l]

https://drive.google.com/drive/folders/1emyaw6aQvhFgFC_RjK1Qo9c1sTRr-avk https://drive.google.com/drive/folders/1W-2EEC5mclFzzWrp65u7JDuLDiaA_vPX 你好，我們在論文中提到的DRCT/HAT/DRCT-L/HAT-L的參數量是參考basicsr的log 我們當初CVPRW的版本中的確實存在錯誤，這是我個人的不慎。 後續我們已更新正確的數值到README以及重新上傳正確的數值到arxiv上，請再參考 https://github.com/ming053l/DRCT 以及正確的arxiv版本 -- 當初FLOPs我們都是用3x64x64的影像當作輸入去計算，(我們當初使用thop) 會造成具體差別的原因我並不清楚是哪個層面影響的，很抱歉我不能回答你 之前我有試著去對DRCT的STL進行改進(例如使用相對複雜的window-attention/加入CAB/採用mamba的Block)，但不同設備訓練起來速度差非常多，不曉得是不是設備的影響 Hello, The parameter counts for DRCT/HAT/DRCT-L/HAT-L mentioned in our paper are based on the logs from BasicSR. There was indeed an error in the CVPRW version of our work, which was due to my personal oversight. We have since corrected the figures and updated them in the README, as well as uploaded the correct values to the arXiv version. Please refer to https://github.com/ming053l/DRCT and the corrected arXiv version for accurate information.

[xduzhangjiayu]

@ming053l 作者您好,
我使用这个工程去训练Real-DRCT-L-GANx4模型，发现保存下来的checkpoints有463MB大小，远远大于模型27M的参数，请问这是为什么呢？
另外我从这个项目的主页的表格来看，DRCT是要比HAT快的，但是从我实测来看HAT是要比DRCT快很多的。

期待您可以解答我的疑惑，谢谢~

[ming053l]

Regarding the complexity data in DRCT's official repo, it's based on the log generated by BasicSR package after training, which might be the main cause of your concern about this. Here's the data comparison for the DRCT-L model: 1. Parameters reported by BasicSR (27,285,295): - This number is calculated through model.parameters() - Only counts parameters that need training (requires_grad=True) - Doesn't include parameters that don't need training (like running mean/variance in batch normalization) - This number represents the actual parameters involved in training 2. params_ema parameters (37,508,911): - This includes all parameters in the model - Includes parameters that don't need training (like batch normalization statistics) - May include some additional buffers - EMA version saves moving averages of all parameters You can refer to the above regarding concerns about file size and model size. As for reading speed, in our experiments, DRCT is faster than HAT. However, due to dense connection requiring storage of all feature map calculation results within RDG, this significantly tests memory capacity and cache access speed. This might be one reason for the difference, but I'm not entirely certain due to equipment inconsistency (which is hard to standardize) - these are more from my own observations. In our paper, we used the following equipment specifications for experiments: CPU: Intel(R) Xeon(R) Gold 6248R CPU @ 3.00GHz GPU: NVIDIA RTX 3090x4 RAM: 442GB When replacing DRCT components, we tried introducing Mamba's VSSM architecture, but training and inference speeds were far below expectations. Later, we saw other researchers integrating Mamba into the DRCT architecture, and it seemed quite fast, which is one basis for my speculation that speed inconsistencies come from equipment differences. This is for your reference. captainzz ***@***.***> 於 2025年1月16日 週四 下午6:18寫道：

…

@ming053l <https://github.com/ming053l> 作者您好, 我使用这个工程去训练Real-DRCT-L-GANx4模型，发现保存下来的checkpoints有463MB大小，远远大于模型27M的参数，请问这是为什么呢？ 另外我从这个项目的主页来看，DRCT是要比HAT快的，但是从我实测来看HAT是要比DRCT快的。 期待您可以解答我的疑惑，谢谢~ — Reply to this email directly, view it on GitHub <#36 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A4P3JL5BBEMORI7WEBR7VCL2K6BQJAVCNFSM6AAAAABUIIQV46VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOJVGEZDEMBYGU> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[WZX2022]

@ming053l 非常感谢您的回复！