COMP 6650: Deep Learning

Spring 2021 / Auburn University

This repository is my implementation of assignments and projects of COMP 6650: Deep Learning. The assignments were from CS231n: Convolutional Neural Networks for Visual Recognition by Stanford University.

Useful Link

Course Page: http://cs231n.stanford.edu/
Github Page: https://cs231n.github.io/
Lecture Videos: YouTube (2017)

Assignment 1

✅ Q1: k-Nearest Neighbor classifier

✅ Q2: Training a Support Vector Machine

✅ Q3: Implement a Softmax classifier

✅ Q4: Two-Layer Neural Network

✅ Q5: Higher Level Representations: Image Features

Assignment 2

✅ Q1: Fully-connected Neural Network

✅ Q2: Batch Normalization

✅ Q3: Dropout

✅ Q4: Convolutional Networks

✅ Q5: PyTorch on CIFAR-10

Version: PyTorch 1.4.0

Projects

Image Classification

In this project, I've implemented several DNN models (Feedforward NN, CNN, RNN, GRU, LSTM and BiLSTM) using PyTorch on FashionMNIST Dataset and compared their performance. The checkpoints can be found here. The comparison is shown below:

Model	Num Layers	Hidden Size	Num Epochs	Learning Rate	Batch Size	Accuracy (Train)	Accuracy (Test)
Feedforward	1	50	10	0.001	64	83.62 %	82.38 %
CNN	2	...	10	0.001	64	88.53 %	87.70 %
RNN	2	256	10	0.001	64	88.06 %	86.37 %
GRU	2	256	10	0.001	64	89.95 %	87.80 %
LSTM	2	256	10	0.001	64	85.40 %	84.06 %
BiLSTM	2	256	10	0.001	64	84.47 %	83.47 %

Image Classification on CIFAR10

In this project, I've finetuned VGG16 for image classification on CIFAR10 dataset. Originally, VGG16 was trained on Imagenet (1000 classes), that's why I have to modify the last layer (to 10 classes) . I also added one additional fully connected layer on top of the original VGG16 model. As I've finetuned the model, only the last 2 layers have been trained keping other layers frozen. The checkpoints can be found here. Hyperparameters are shown below: