cydiachen / fsrnet_pytorch Goto Github PK

Dear Friend, Having tried several ways, I tried to reproduce the performance of FSRNet by using Pytorch. I am Now transferring to another new project FSRNet, You can download the code and run at your own machine. Feel free to contact me.

Python 100.00%

fsrnet_pytorch's Introduction

FSRNet Pytorch

Dear friends, Thank you for keep tracking in this implementation of FSRNet (CVPR 2018 Oral Paper)

I have been spent the whole summer as an intern in Iluvatar.ai. I have been back to school, so I have time to complete the Project.

I rewrite the Train.py and other model code completely. Now I am uploading pretrained Weights on BaiduNetDisk together with Training Set.

Based on WaveletSRNet, I altered the code by adopting FSRNet network structure.

Prerequisites

Python 3.6
Pytorch 1.0 or newer (Pytorch > 0.4 should be ok)
matplotlib
skimage

Train

Change the option in Train.py to set the dataset's directory. I am using CelebAHQ-MASK as the training set. The GroundTruth is generated by zllrunning/face-parsing.PyTorch(https://github.com/zllrunning/face-parsing.PyTorch) with pretrained model.

Dataset Link: https://pan.baidu.com/s/1HEECUyKI5GOSrd7NPlm-ow 密码:z2ud

Test

ON GOING :| PYTHON AND NOTEBOOK WILL BE PROVIDED. Pretrained Weights:链接:https://pan.baidu.com/s/1ZkgABGefsMjO6XhhvlBzRA 密码:libl

Result

Citation

If you find FSRNet useful in your research, please consider citing (* indicates equal contributions):

@inproceedings{CT-FSRNet-2018,
  title={FSRNet: End-to-End Learning Face Super-Resolution with Facial Priors},
  author={Chen, Yu* and Tai, Ying* and Liu, Xiaoming and Shen, Chunhua and Yang, Jian },
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2018}
}

fsrnet_pytorch's People

Contributors

Stargazers

Watchers

Forkers

nmber5 tomdkim shaohuawan soccergame llltttppp liuguoyou tony109060581 snowbrain mingyang96ong yanghedada jimmy0723123 zt706 anperlanch

fsrnet_pytorch's Issues

models

缺少文件models.py,运行train.py时显示缺少models,我使用之前版本的models.py时显示缺少losses

关于 CelebAHQ-MASK大小的问题

请问在训练之前需要将图片尺寸剪裁至128x128大小吗文件夹data\CelebA-HQ-img里的Parsing_Maps也需要同样的处理吗谢谢

关于复现结果上不来，我观察到一点问题

网络模型用于训练的图像是经过切分对齐后的，但是heatmaps标签是未经过对齐和切分的，不准确的heatmaps会扰乱fineSR的训练过程。作者再复现时用的什么样的heatmaps标签，大家一起讨论一下，看是不是我搞错了？
输入HR图像（128 * 128）：

heatmaps标签中的第一张（64 * 64）：

我认为这个heatmaps需要对齐和切分处理

how to generate fsrnet lr img ?

how to generate lr img, can you share a detail steps in fsrnet?

Questions about training details

Thank you for your work very much.I have some questions about training,can you please tell me?
1.Will the parameters in PriorEstimationNetwork change or not?
2.Have the images of training prior network and training fsrnet normalized?
Thank you for your help.

Is this the full implementation?

I was confused by reading the description of this repo and the introduction in README file. Is this now the full implementation of FSRNet? Does this reproduce the results in the original paper?

关于代码的问题

你好，我最近在做人脸超分的任务，以fsrnet为基础，我进行了一些网络结构的改进，所得的结果也比fsrnet要好一些，但是都达不到论文里的结果。请问可以对我做出一些指导或者建议吗？万分感谢！

Steps to follow to run this projects ?

Can someone explain to me what are the steps to follow for executing this projects : first file to execute, second one ...?
I want to execute it under Google Colab.
Thanks.

问一下大家目前尝试的最好的结果是什么了

我看实现的最好的结果PSNR在22-23之间，有没有更好一些的了，或者实现出原论文的效果的了？

您好!首先感谢您的工作，另外，请问.npy的文件可以分享一下嘛

opt = parser.parse_args()

运行train.py报错SystemExit2 请问这个怎么解决

Unable to download pretrained model

Hello would it be possible to provide an alternative download link to the pretrained model.
I am unable to register a baidu account Because I do not live in China.

Machine translation / 机器翻译 :

您好，有可能提供预训练模型的替代下载链接。
我无法注册百度帐户，因为我不居住在**。