Convolutional Neural Network and Contrastive Autoencoder on EMNIST

In this project, we are going to evaluate the performance of convolutional neural network (CNN) and contrastive autoencoder (CAE) models by conducting empirical study on simple image data (EMNIST dataset) [1]. This dataset consists of 28x28 images of handwritten characters that belong to 47 classes.

[1] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre van Schaik. EMNIST: an extension of MNIST to handwritten letters. arXiv preprint arXiv:1702.05373, 2017.

Run the code

Package required

• Python 3.5 (or later)
• Tensorflow (https://www.tensorflow.org/)

Parameters

• lr: initial learning rate
• mm: momentum
• bsz: batch size

CNN

Run the code

Train CNN

python --task="train_cnn" --lr=0.1 --mm=0.2 --bsz=32

Cross Validation CNN

python --task="cross_valid_cnn"

Test CNN

python --task="test_cnn" --lr=0.1 --mm=0.2 --bsz=32

Autoencoder

Run the code

Train AE

python --task="train_ae" --lr=0.1 --mm=0.2 --bsz=32

Cross Validation AE

python --task="cross_valid_ae"

Test AE

python --task="evaluate_ae" --lr=0.1 --mm=0.2 --bsz=32

Note

COMP5212 - Machine Learning Programming Assignment 2 in HKUST