An open API service indexing awesome lists of open source software.

https://github.com/dgedon/cyclegan-for-emojis

CycleGAN implementation for Emojis
https://github.com/dgedon/cyclegan-for-emojis

cyclegan gan pytorch reimplementation

Last synced: 13 days ago
JSON representation

CycleGAN implementation for Emojis

Awesome Lists containing this project

README

          

# CycleGAN for Emoji Style Transfer

This is the repository for the project of module 3 in the [WASP](https://wasp-sweden.org/graduate-school/)
graduate school course [Deep Learning and GANs](https://wasp-sweden.org/graduate-school/ai-graduate-school-courses/) by
[Sofia Ek](http://www.it.uu.se/katalog/sofli286), [Carmen Lee](https://www.it.uu.se/katalog/carle978) and [Daniel Gedon](https://www.it.uu.se/katalog/dange246).

We are reimplementing the CycleGAN paper by Zhu et al, see [here](https://junyanz.github.io/CycleGAN/). We are taking the
same setup as in assignment 4 of the course by Roger Grosse, see [here](http://www.cs.toronto.edu/~rgrosse/courses/csc321_2018/).
That means that we take the accompanying dataset, containing emojis from Apple and Windwos and try to transfer the style between them.
The learnt transfer can be seen in the image below.

## Results

The following restults are obtained after training for 220 epochs from scratch.

![](doc/epoch220-app.png?raw=true)
*Transfer from Apple style emojis to Windows style ones*

![](doc/epoch220-win.png?raw=true "Test")
*Transfer from Windows style emojis to Windows style ones*

## Code Usage

To run the code with standard settings use
```bash
python main.py
```
The results above are obtained with
```bash
python --batch_size=16 --max_epochs=500 --learning_rate=3e-4 --beta1=0.5 --beta2=0.999 --loss_lambda=2
```