python 多帧 超分辨_Pytorch实现多帧超分辨率(MFSR)网络HighRes-net

HighRes-net: Multi Frame Super-Resolution by Recursive Fusion

Pytorch implementation of HighRes-net, a neural network for multi frame super-resolution (MFSR), trained and tested on the European Space Agency's Kelvin competition.

Computer, enhance please!

source: ElementAI blog post Computer, enhance please!

A recipe to enhance the vision of the ESA satellite Proba-V

Hardware:

The default config should work on a machine with:

GPU: Nvidia Tesla v100, memory 32G

Driver version: CUDA 10.0

CPU: memory 8G to enable running jupyter notebook server and tensorboard server

If your available GPU memory is less than 32G, try following to reduce the memory usage

(1) Work with smaller batches (batch_size in config.json)

(2) Work with less low-res views (n_views and min_L in config.json, min_L is minimum number of views (n_views))

According to our experiments, we estimated the memory consumption (in GB) given batch_size and n_views

batch_size \ n_views and min_L

32

16

4

32

27

15

6

16

15

8

4

0. Setup python environment

Setup a python environment and install dependencies, we need python version >= 3.6.8

pip install -r requirements.txt

1. Load data and save clearance

Download the data from the Kelvin Competition and unzip it under data/

Run the save_clearance script to precompute clearance scores for low-res views

python src/save_clearance.py --prefix /path/to/ESA_data

2. Train model and view logs (with TensorboardX)

Train a model with default config

python src/train.py --config config/config.json

View training logs with tensorboardX

tensorboard --logdir='tb_logs/'

3. Test model

Open jupyter notebook and run notebooks/test_model.ipynb

We assume the jupyter notebook server runs in project root directory. If you start it in somewhere else, please change the file path in notebooks accordingly

You could also use docker-compose file to start jypyter notebook and tensorboard

Authors

License

This repo is under apache-2.0 and no harm license, please refer our license file

Acknowledgments

Special thanks to Laure Delisle, Grace Kiser, Alexandre Lacoste, Yoshua Bengio, Peter Henderson, Manon Gruaz, Morgan Guegan and Santiago Salcido for their support.

We are grateful to Marcus Märtens, Dario Izzo, Andrej Krzic and Daniel Cox from the Advanced Concept Team of the ESA for organizing this competition and assembling the dataset — we hope our solution will contribute to your vision for scalable environmental monitoring.