amitadate / S-LSTM-GAN-MNIST

This repository contains the source for the paper "S-LSTM-GAN: Shared recurrent neural networks with adversarial training"

LSTM-GAN-MNIST

Link to the paper, visit: Link

To easily cite our Paper, visit : Link

Brief description:

Using the MNIST set to experiment with GANs using LSTM's

Model

Following a generic generative adversarial network, the model consists two networks trained in parallel, and sharing weights. The pink portion of the model is the generator and the orange-brown portion is the discriminator. For purposes of clarity the image is split into quadrants here, but in other experiments the attempt was to split the image into pixels in an attempt to create a generator that could create digits pixel by pixel using long range memory. Up to now the best results have occurred with splitting the image into 16 sections, beyond that the model fails.

Generator

Discriminator

Experiments

TIMESTEP MODEL

Variable	Value
timesteps	4
lstm_layers_RNN_g	6
lstm_layers_RNN_d	2
hidden_size_RNN_g	600
hidden_size_RNN_d	400
lr	1e-4
iterations	> 2.5e6

SAMPLES

0	1	2	3	4	5	6	7	8	9

TIMESTEP MODEL

Variable	Value
timesteps	16
lstm_layers_RNN_g	6
lstm_layers_RNN_d	2
hidden_size_RNN_g	600
hidden_size_RNN_d	400
lr	2e-4:GEN/1e-4:DISC
iterations	> 5e5

SAMPLES

0	1	2	3	4	5	6	7	8	9