PixelArtAI

Thanks for your porject! I have a quesention about download_and_convert_data.py line 255 num_shards = int(len(imgs) / 2500) what 2500 means? If I use others dataset,how can I modify it?

训练时，同一个网络，网络较小时，可以正常训练，网络较大时，损失函数很大，不能正常训练。

1

我是用多尺度的残差网络训练去噪任务，当参数设置比较小，网络大概30层时，初始loss很小，可以正常训练。但是当网络增加到大概300层时，训练的初始loss非常非常大，网络无法训练，不收敛。 其他尝试： 1.尝试了几个其他的网络，都有这个情况。大网络不能正常训练。 2.尝试了调整lr，网络权重初始化。也不能解决。 请问有什么解决思路吗，或者有什么可能的原因吗？

How about Inference time

1

Thanks for sharing such a great work, the inference time is also a point to be considered for lightweight. How does the inference time compare to IMDN?

PSNR and SSIM of DIV2K only

1

Hi, thanks for you share, but I have a question of the result. Have you ever trained alone with D and what was the result of Set5 and Set14? Thanks.

[Bug] RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB

3

### Prerequisite - [X] I have searched [Issues](https://github.com/open-mmlab/mmediting/issues) and [Discussions](https://github.com/open-mmlab/mmediting/discussions) but cannot get the expected help. - [X] I have read the [FAQ documentation](https://mmediting.readthedocs.io/en/1.x/notes/4_faq.html) but cannot get the expected help....

Need pixel loss training first?

1

Thanks for open source, Does it needed to use pixel loss training first, then load the pre-training model and add GAN loss to continue training?

Will the dataset be open source？

1

Great worker，will the dataset be open source？