You are on page 1of 1

AI3001

Deep Learning Course


Assignment-2a
Instructor: Dr. Anupam Sobti

Instructions for Submission:

1. Assignment is due at 11:59:59 PM on Tuesday, Mar 5, 2024.


2. Please follow the course policies.
3. You need to submit Python code in .ipynb notebook with the output clearly defined.

4. Please use WANDB to log loss values and hyperparameters for various experiments.
5. Please start early for this assignment.

Question 1. In this assignment, you will train a generative adversarial network on the MNIST dataset and
learn how to generate images that resemble the digits from the MNIST dataset.

(a) Propose two architectures for generating the digits in the MNIST dataset provided. Ar-
chitecture 1 should use purely linear layers and Architecture 2 should use purely convo-
lutional/pooling layers. Explain the rationale behind choosing this architecture.
(b) What would you use for discriminator and generator loss functions? Modify this to
Wassertein GAN loss function and report the changes.
(c) For the min-max game for optimization, report methods used to ensure network conver-
gence, specifically, comment on the learning rate schedule and the number of iterations
required for the “min” and “max” phases of training.
(d) For the best GAN generated so far, add the Ambiguous MNIST dataset in your training
data.

You might also like