PerceptualLossNetwork

This repository is for CS577 (Deep Learning) @ IIT The work in this repository is based on the paper:

Perceptual Losses for Real-Time Style Transfer and Super Resolution

@article{
title={Perceptual Losses for Real-Time Style Transfer and Super Resolution},
author={Justin Johnson, Alexandre Alahi, Li Fei-Fei},
publisher={Dept. of Computer Science, Stanford University},
year={2016}
}

This implementation was inspired from the offical pytorch example from facebook research:

Installation

MacOS (CPU)

• Install conda
• Set env by either:
  • Create conda enviornment: conda create --name <env> --file macOS_local_req.txt
  • Install into existing conda environment: conda install -n <env_name> macOS_local_req.txt

Commands:

Example to test the model ( stylize an image ) :

python3 src/stylize.py --image data/test/COCO_train2014_000000363111.jpg --model data/trained_models/mosaic_id5320.pth

To train a model from scratch

python3 src/trainer.py --data-dir data/training_data/ --style data/style/mosaic.jpeg

To train with checkpoints

python3 src/trainer.py --data-dir data/training_data/ --style data/style/mosaic.jpeg --save 1 --checkpoints-path data/dir_to_save

To retrain a model

python3 src/trainer.py --data-dir data/training_data/ --style data/style/mosaic.jpeg --retrain 1 --retrain-model data/trained_models/tree_to_retrain.pth

To visualize the losses

python3 src/visualize_losses.py

To see the different arguments of trainer.py:

python3 src/trainer.py

Perceptual Loss

Transformation Network Architecture

VGG16 Architecture:

Results