Super Resolution Model and Web App

This repository contains a super resolution network upscaling images of a factor of 4 and a web app. The final web app looks like th is:

Data

• data from kaggle.com
• raw data saved in data/HR (high resolution) and data/LR (low resolution)
• 100 images
• data exploration: notebooks/data-exploration.ipynb

Preprocessing

• preprocessing.py
  • splits the data in validation and training
  • saves splits in h5 files to data/train.h5 and data/valid.h5

Model

• train 2 models:
  • SRResNet
  • SRGAN (ToDo)
• based on: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, Ledig et al. (2017), https://doi.org/10.48550/arXiv.1609.04802
• model code based on https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Super-Resolution
• converted to pytorch lightning
• added NNI for hyperparamtertuning
• added mlflow

Training

SRResNet

• run script with desired parameters, e.g. python3 srresnet.py --n-blocks 8 --batch-size 12 --n-channels 22 --gpus 1 --n-epochs 200
• use batch script saved in jobs/start_job.sh or start_job_nni.sh to include hyperparameter tuning using nni
• results are not reported to nni, but to mlflow
• example files for hyperparamter tuning are in nni/config/config.py and nni/search_space/search_space.json

Evaluation

• example plots are in notebooks/plot-predictions.ipynb

Deployment

• model is served using torchserve
• the handler script is based on the BaseHandler class and can be found in deployment/srnet_handler.py
• create the .mar file using the create-mar.sh script
• run docker build -t srnet-mar:v1 . to create the docker image
• docker run -p 8080:8080 -p 8081:8081 srnet-mar:v1
• check registeres models: curl http://127.0.0.1:8081/models
• make predictions: curl http://127.0.0.1:8080/predictions/srnet -T ../data/LR/0.png
• The config.properties file is needed to change the maximun output size

Web App

• simple web app is stored in folder app
• run docker-compose up to start the services
• navigate to localhost:9696 in browser