Style transfer with catalyst

This repository shows experiment in realtime style transfer with catalyst deep learning framework. The experiment is based on article Perceptual Losses for Real-Time Style Transfer and Super-Resolution

Examples

Usage

1. Install dependencies: pip3 install torch==1.6.0 catalyst==20.8.2 numpy tensorflow==2.0.0 tensorboard

   Attention: Catalyst don't have guaranteed backward compatibility, please use only specified version.

2. Configure experiment:
   Fill all fields marked by "{SPECIFY}" tag in config.yml

3. Run training:
   catalyst-dl run --config config.yml --verbose
   Well recommendation is to install https://github.com/NVIDIA/apex and use more than one GPU. Catalyst provides multi-GPU and distributed training "out-the-box"

4. Inference:
   See infer_catalyst.py to find details.

5. Ready models:
   You can find weights for 512x512 ImageTransformer in ./models