Neural Style Transfer in FastAI

Implementing "Perceptual Losses for Real-Time Style Transfer and Super-Resolution" with FastAI framework. Apply hook (PyTorch mechanism) to calculate loss.

Style Source

An Artwork from HK Artist, Cuson Lo

Stylized Samples

Reference

1. fastai -- Extracting Intermediate Features Using Forward Hook
2. fastai -- 06_cuda_cnn_hooks_init.ipynb, Deep Learning for Coder part 2 v3
3. fastai -- documentation on what a callback can unpack from kwargs
4. fastai2 -- another implementation on fast neural style transfer
5. pytorch -- fast-neural-style
6. cs231n -- total variation loss implementation
7. tensorflow -- fast neural style transfer with a rich documentation
8. arxiv -- Instance Normalization: The Missing Ingredient for Fast Stylization
9. arxiv -- Perceptual Losses for Real-Time Style Transfer and Super-Resolution
10. medium -- Practical Techniques for getting Style Transfer to Work

Log

[05/04/2020]

• completed hooks callback and tensorboard callbacks

[08/04/2020] pending experiments:

• try transformer with "downsample-first-upsample-final" design (for speeding up inference, training)
• try total variation loss
• try globally reducing weight for all losses
• place test sets into validation dataloader (code refactoring)