My first-ever rendition of Generative Adversarial Networks (GANs) which are a deep-learning-based generative model. GANs learn to generate new data with the same statistics as the training set. It consists of two networks, the Generator & and the Discriminator, playing an adversarial game against each other. The Generator produces fake data to trick the Discriminator whereas the Discriminator inspects the fake data & and determines whether it is real or fake. They are randomly initialised in the beginning & and are trained simultaneously.
I ran it for 100 Epochs using the MNIST dataset, using the Adam Optimizer. The following shows a series of images produced by the generator as it was trained.
The images begin as random noise, and increasingly resemble hand written digits similar to the MNIST dataset over time.
The Adaptive Moment Estimation (Adam) Optimizer & Stochastic Gradient Descent (SGD) are optimization techniques used to improve the model's performance. Below shows the data generated after 30 Epochs using both optimizers:
Given how the number of Epochs is relatively small, I decided to go with with the Adam Optimizer (Adam) as it converged faster compared to the SGD Optimizer during my testing. There are more optimization algorithms available in the torch.optim
package, which could have performed better & I suggest trying these out as well.
-
Get a Google Account here to use Google Colab.
-
Test it using this Google Colab Notebook.
Note that the maximum number of Epochs is set to 40 in the notebook. Using a T4 GPU, it would take around 14 minutes for it run completely. The output from a max_epochs value of 40 has been included the notebook for easy viewing.