nathan-yan / mnist-neural-automata

Generating novel MNIST digits using neural cellular automata.

Growing MNIST digits with differential cellular automata.

This is based heavily on Mordvintsev et al. [https://distill.pub/2020/growing-ca/]. To generate novel digits you can run automata_test.py, which will use the pretrained models inside the ./models directory.

To train your own network, run automata.py first to train the differential cellular automata and autoencoder to generate good looking MNIST digits. Then using the generated model, run automata_persistent_mnist.py to train the automata to generate persistent digits. Finally using the new generated model, run gan_mnist.py to train a Generative Adversarial Network on the latent vectors of the autoencoder. Finally using that model and the model from automata_persistent_mnist.py, you can run automata_test.py. The code right now is kinda messy so I'm working on cleaning it up and making the training pipeline easier.

Using this autoencoder/GAN technique is definitely not the best way of approaching it, and I'm hoping to train the whole thing end-to-end as a GAN to generate more realistic looking digits.

Dependencies

• numpy>=1.16.0
• matplotlib>=3.0.0
• torch>=1.4.0

mnist-neural-automata's Issues

How is this run?

Is it:
python gen_mnist.py

?

what is the difference between processed and raw data

Hi

I found this model very interesting. And when I tried to run, I found that the results trained from processed data and raw data had dramastic difference. What's the difference?