bytedance / rlfn Goto Github PK

Winner of runtime track in NTIRE 2022 challenge on Efficient Super-Resolution

License: Apache License 2.0

Python 100.00%

lightweight super-resolution ntire22challenge

rlfn's Introduction

Residual Local Feature Network

Our team (ByteESR) won the first place in Runtime Track (Main Track) and the second place in Overall Performance Track (Sub-Track 2) of NTIRE 2022 Efficient Super-Resolution Challenge.

_model	_Runtime[ms]	_Params[M]	_Flops[G]	_Acts[M]	_{GPU Mem[M]}
RLFN_ntire	27.11	0.317	19.70	80.05	377.91

Open-Source

For commercial reasons, we don't release training code temporarily, please refer to EDSR framework and our paper for details.

Paper of our method [arXiv]
Report of our performance [NTIRE22 official report]
The pretrained model and test code in challenge.

Testing

We modified the official test code. To reproduce our result in the ESR challenge, please install PyTorch >= 1.5.0.

run python test_demo.py to generate image results.
All test results will be saved in the folder data/DIV2K_test_LR_results

rlfn's People

Contributors

Stargazers

Watchers

rlfn's Issues

ESA中的分组卷积和代码对应不上的问题

从代码中看ESA中分组卷积的代码如下所示：
self.conv3 = conv(f, f, kernel_size=3, padding=1)
请问以哪个为准～

about Contrastive Loss

Hi, thanks for your great work!

I have a question about contrastive loss.
What does (Ypos, Yneg, Yanchor) stand for?

More Pretrain Model Need!

rlfn_ntrie_x2.pth need, please. 谢谢

Question

How to visualize the extracted features?

conv_f 在ESA strcture中的作用

Hi:
请问下model/rlfn_ntire.py中class ESA中的self.conv_f的作用是什么，在论文中Figure 3的ESA的结构中没有这个卷积层？

Pre-trained models

First of all, thank you for sharing excellent work!!

It seems that only the pre-trained model with scale factor 4 is provided.

Can you share other pre-trained models with different scaling factors including x2?

How many iterations of the model can you get the best performance in the benchmark？

Benchmark like Set5, Set14 and so on.
According to the paper, the initial learning rate is 5e-4, and is halved every 2e5 iterations, but the total number of iterations is not specified.

New Super-Resolution Benchmarks

Hello,

MSU Graphics & Media Lab Video Group has recently launched two new Super-Resolution Benchmarks.

Video Upscalers Benchmark: Quality Enhancement determines the best upscaling methods for increasing video resolution and improving visual quality.
Super-Resolution for Video Compression benchmark aims to test Super-Resolution methods on compressed videos and select the best model for each video codec standard.

If you are interested in participating, you can add your algorithm following the submission steps:

We would be grateful for your feedback on our work!