XyChen
@wyuzl DF2K下载和处理可以参考[BasicSR](https://github.com/XPixelGroup/BasicSR/blob/master/docs/DatasetPreparation.md#image-super-resolution)。ImageNet直接从官网下载后参考上述链接制作meta_info文件即可。
@wyuzl 数据增强在代码中有体现，具体实现你可以参考源码[paired_image_dataset.py](https://github.com/XPixelGroup/BasicSR/blob/master/basicsr/data/paired_image_dataset.py)。提供的训练option文件里默认是打开数据增强的，不需要额外设置。数据集本身并没有做增强处理，按照上个回复提供的链接制作就行。制作好的DF2K sub-images体积过大，不方便上传。
@wyuzl 检查了一下确实是3450。这里是错算成了2560+800=3360...
@HUSTLX Has the problem been solved? The training log is not easily feasible but I can tell you the training process is really stable.
how long have you spent when you train ImageNet ? and how many GPUS you used ? thank you very much !
@weibiansanjue01 About 4 days for HAT and about 8 days for HAT-L on 8 A100 GPUs.
If you do not use the tile mode, the full resolution will be calculated when testing.
@EvgeneyZ Thanks for your invitaion and I really appreciate your great work about the benchmarks. I consider participating in the project when I have time to do it.
Hi @ichsan2895, I haven't integrated tile mode into this repo. I will add this function soon. Before that, you can try our method in the [replicate](https://replicate.com/xpixelgroup/hat).
The tile mode is provided for limited GPU memory when testing. The setting can be referred to https://github.com/XPixelGroup/HAT/blob/1d157411076a2d485b13c711a6a521abf62758b0/options/test/HAT_tile_example.yml#L7-L9
@ziippy The tile mode will indeed cause slight performance change. The smaller the image size and the smaller the tile size, the greater the performance change. When your GPU can...