jacobgil / pytorch-zssr Goto Github PK

PyTorch implementation of 1712.06087 "Zero-Shot" Super-Resolution using Deep Internal Learning

License: Apache License 2.0

Python 100.00%

pytorch deep-learning super-resolution computer-vision

pytorch-zssr's Introduction

Unofficial PyTorch implementation of "Zero-Shot" Super-Resolution using Deep Internal Learning

Unofficial Implementation of 1712.06087 "Zero-Shot" Super-Resolution using Deep Internal Learning by Assaf Shocher, Nadav Cohen, Michal Irani.

Official Project page: http://www.wisdom.weizmann.ac.il/~vision/zssr/

Paper: https://arxiv.org/abs/1712.06087

This trains a deep neural network to perform super resolution using a single image.

The network is not trained on additional images, and only uses information from within the target image. Pairs of high resolution and low resolution patches are sampled from the image, and the network fits their difference.

TODO:

Implement additional augmentation using the "Geometric self ensemble" mentioned in the paper.
Implement gradual increase of the super resolution factor as described in the paper.
Support for arbitrary kernel estimation and sampling with arbitrary kernels. The current implementation interpolates the images bicubic interpolation.

Deviations from paper:

Instead of fitting the loss and analyzing it's standard deviation, the network is trained for a constant number of batches. The learning rate shrinks x10 every 10,000 iterations.

Usage

Example: python train.py --img img.png

usage: train.py [-h] [--num_batches NUM_BATCHES] [--crop CROP] [--lr LR]
                [--factor FACTOR] [--img IMG]

optional arguments:
  -h, --help            show this help message and exit
  --num_batches NUM_BATCHES
                        Number of batches to run
  --crop CROP           Random crop size
  --lr LR               Base learning rate for Adam
  --factor FACTOR       Interpolation factor.
  --img IMG             Path to input img

pytorch-zssr's People

Contributors

Stargazers

Watchers

pytorch-zssr's Issues

IndexError: too many indices for array on line 46 in train, cpu_loss = error.data.cpu().numpy()[0]

(base) C:\Users\Luke Chen\Desktop\pytorch-zssr-master>python train.py --crop 30 --num_batches 8 --img img.png
train.py:18: UserWarning: nn.init.kaiming_normal is now deprecated in favor of nn.init.kaiming_normal_.
  init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
  0%|                                                                                            | 0/8 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "train.py", line 118, in <module>
    train(model, img, args.factor, args.num_batches, args.lr, args.crop)
  File "train.py", line 46, in train
    cpu_loss = error.data.cpu().numpy()[0]
IndexError: too many indices for array

FileNotFoundError: [Errno 2] No such file or directory: 'lincoln.png'

(torch14) C:\Users\15479\Desktop\pytorch-zssr-master>python train.py --img lincoln.png
Traceback (most recent call last):
File "train.py", line 105, in
img = PIL.Image.open(args.img)
File "E:\anaconda\envs\torch14\lib\site-packages\PIL\Image.py", line 2891, in open
fp = builtins.open(filename, "rb")
FileNotFoundError: [Errno 2] No such file or directory: 'lincoln.png'

I'm a green hand.Please tell me What should be after --img？
How to set？
Please ，help me bro！！！

compare with the original txt

I use the defalut parameters.
lr = 1e--5
nums = 15000
in fact, I test in Set14，and get the results as:
x2,psnr:29.75,ssim:0.88
however, the original txt is:
x2,psnr:33.00,ssim:0.9108.
I want to know ,did you test in Set14,and what's your results.

Kernel estimation?

I was wondering if you had a chance to look into the kernel estimation part of the paper, as that is what really lets ZSSR be more useful than other SR methods (per the paper). If not, do you have any guidance for trying to implement it? In particular, how can image resizing with an arbitrary kernel be done?

hello，I want to know zzsr use how much training pairs and whether it need 16000 epoches?

thanks

CUDA out of memory.

I ran everything with default settings just to test the code ( run train.py --img lincoln.png).
I got the 2x enlarged image.

Then I used this 2x image as input to magnify it to 4x. During this run, all the 15000 epochs were completed and a new image file called 'low_res.png' was found in the folder. But after this, the program stopped without completion with the following message :

CUDA out of memory. Tried to allocate 788.00 MiB (GPU 0; 4.00 GiB total capacity; 2.38 GiB already allocated; 531.39 MiB free; 128.68 MiB cached)

Also, during the running of the program, I checked the task manager and it was always like CPU(near 100%), GPU(2%).

Can someone help me understand the issue and help me find a solution?

jacobgil / pytorch-zssr Goto Github PK

pytorch-zssr's Introduction

Unofficial PyTorch implementation of "Zero-Shot" Super-Resolution using Deep Internal Learning

Usage

pytorch-zssr's People

Contributors

Stargazers

Watchers

Forkers

pytorch-zssr's Issues

Recommend Projects

Recommend Topics

Recommend Org