Skip to content

Latest commit

 

History

History
84 lines (56 loc) · 3.67 KB

README.md

File metadata and controls

84 lines (56 loc) · 3.67 KB
Raw

trump_make

Trump Fakes

An explanation of how DCGAN works, can be found here:

https://medium.com/@keisukeumezawa/dcgan-generate-the-images-with-deep-convolutinal-gan-55edf947c34b

PyTorch has also released a turorial: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html

Similarly, code that runs the same way can be found here: https://debuggercafe.com/implementing-deep-convolutional-gan-with-pytorch/

An experiment in DCGAN using Donald Trump's face as the dataset benchmark. Feel free to change the parameters, add or remove Trump faces, etc. This is more of a game to understand how GAN (and specificially Convolutional GAN) work, and what are their challenges and results (aka deep fakes).

The data is under (suprise!) data/ and was extracted from kaggle using the script preprocess.py. The original dataset was found in https://www.kaggle.com/mbkinaci/trump-photos/data and the script will resize pics, detect faces, extract them as pics and copy them.

Some removal is also needed (as some pics are too small, or some contain non-trump faces), or you can use the data/trump_faces` as is.

The dcgan.py has params and hyper-params, and you can play freely with it, getting different results. This is a game between two networks:

  • The Generator (aka G) will try to generate better and better fakes
  • The Discriminator (aka D) will try to detect fakes as best as possible
  • When they reach equilibrium, e.g., when G generates good fakes which D can't tell apart from reals, then you should be happy. I haven't gotten there yet.

You can edit the G class as I've added LeakyReLU instead of ReLU and the complexity of the network may need to be increased for better results.

Training with ADAM seems to work, SGD provided worse results for some reason (I usually get better results with SGD in image classification).

Please note the code is taken from the DCGANN example from PyTorch with minor modifications to suit my needs.

At some point I plan on adding tensorboard support instead of matplotlib and mayb trying slightly different achitectures for the G network.

If you fancy, send me a pr and if everything is OK I'll accept it.

UPDATE for Hacktoberfest 2022

First of all, read CONTRIBUTING.md. Then note the following;

  1. I've made a move from pytorch-only code to pytorch-lightining since it abstracts and moves a lot of stuff into classes.
  2. This may have broken functionality. If it did, sorry!
  3. I'm using CUDA 11.6 so torch and torchvision need to match it. If your installation is older, change your virtual environment, but don't commit the changes back
  4. Use PEP8 or keep code tidy please, and document it.
  5. Use python3.9 and pip to setup a virtual environment and run the code on CUDA/GPU.
  6. Use tensorboard to visualise progress, metrics and produced images!

Have Phunz!

UPDATE of October 2024

I've tried very hard to make the DCGAN work. It used to generate mediocre outputs, and then somehow it broke. Back when it was released, there were afew people dicussing it. Below are useful links that may help you if you decide to play with this architecture. Needless to say it has since been surpased.

The original post: https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html?highlight=dcgan

A very good discussion on DCGANS: pytorch/examples#70

And last but not least, some very good hints and tips: https://github.com/soumith/ganhacks

Ultimately I just decided to use a ready made model from lighting bolts: https://pytorch-lightning-bolts.readthedocs.io/en/stable/models/gans.html#dcgan

Yes, I am aware this defeats the purpose, but I got defeated by DCGANs so I gave up.