liamma/ift6135-assignment-1

IFT6135-Assignment

Jin Dong: [email protected]; Liheng Ma: [email protected].

This is the assignment codebase for the IFT-6135: Learning Representation (Deep Learning) class taught by Prof Aaron Courville at UdeM.

Assignment 1: CNN

1. Implement MLP with vectorized back-propogation using Numpy.
2. Implement a standard CNN for MNIST.
3. Implement an improved CNN for MNIST on Kaggle using DL library (we used MxNet).

Assignment 2: RNN

1. Implement a RNN, GRU Cell using PyTorch.
2. Implement the multi-head self-attention module of Transformer.
3. Do experiments on various setting.

Assignment 3: Generative model

1. Build a discriminator to approximate Wasserstein Distance, Jensen–Shannon divergence, etc.
2. Build a Variational Auto-encoder with binary-cross-entropy loss to generate binarized MNIST images.
3. Build a Variational Auto-encoder with MSE loss to generate SVHN images.
4. Build a Wasserstein GAN with gradient-penalty to generate SVHN images.
5. Qualitative and quantative analyze the performance of VAE and GAN.