This repo contains some experiments with Generative Adversarial Networks which could be tested on Google Colab.
Generative Adversarial Networks with convolutional Discriminator and transposed convolutional Generator trainned on Google Colab's GPUs for 490 minutes (~8 hours) and 600000 learning iterations to generate MNIST handwritten digits. At last iteration mean BCE of Discriminator was 0.01261 and mean BCE of Generator was 5.65304. Generator creates images from 100-dimentional random vector. You may find pritrainned model here: latest model checkpoint.
