Background

This repo contains links to University of Michigan EECS 498.008/598.008 Deep Learning for Computer Vision course assignments completed by Elliot using Google Colab, Jupyter Notebook, Python, and PyTorch.

UMich EECS 498.008/598.008 references and builds upon the Stanford University CS231n Deep Learning for Computer Vision course, with additional topics on emerging Computer Vision research developments.

From the UMich course description:
"This course is a deep dive into details of neural-network based deep learning methods for computer vision. During this course, students will learn to implement, train and debug their own neural networks and gain a detailed understanding of cutting-edge research in computer vision. We will cover learning algorithms, neural network architectures, and practical engineering tricks for training and fine-tuning networks for visual recognition tasks."

UMich EECS 498.008/598.008 Course
Stanford CS231n Course

Table of Contents

1. Assignment 1: PyTorch 101; k-Nearest Neighbor Classifier
2. Assignment 2: Linear Classifiers; Two-Layer Neural Network
3. Assignment 3: Fully-Connected Neural Network; Convolutional Neural Network
4. Assignment 4: Object Detection
5. Assignment 5: Image Captioning with Recurrent Neural Networks; Transformers
6. Assignment 6: Variational Autoencoder; Generative Adversairal Networks, Network Visualization, Style Transfer

Assignment 1: PyTorch 101; kNearest Neighbor Classifier

Status: Completed Jan 2024
Link to GitHub Repo (Please request access to this private Repo)

Implementation of these classes and functions:

• KnnClassifier for image data with dimensions (num_images, channels, height, width)
• k-folds cross validation
• compute_distances using nested loops, single loop, and vectorized implementation

Key concepts and skills used:

• PyTorch tensors manipulation
• Euclidean distance calculations

Trained K-Nearest Neighbor classifier for CIFAR-10 image dataset achieved test set accuracy: 33.86%

Assignment 2: Linear Classifiers; Two-Layer Neural Network

Status: Completed Jan 2024
Link to GitHub Repo (Please request access to this private Repo)

Implementation of these classes and functions:

• LinearClassifier for flattened image data in tensors (num_data, flattened_dimensions)
• SVM Loss function and Softmax (Cross Entropy) Loss function forward and backward pass
• TwoLayerNet fully connected neural network for flattened image data
• nn_forward_pass, nn_forward_backward
• nn_train using Stochastic Gradient Descent (SGD)
• Hyperparameters search using grid search

Key concepts and skills used:

• Devise Computational Graphs to create Forward Pass and Backward Backpropagation calculations for addition, multiplication, element selection operations
• Naive Implementations using Loops to understand element-by-element calculations
• Vectorized Implementations to harness the speed-up of vectorized operations
• L2 Weight Regularization

Trained TwoLayerNet model for CIFAR-10 image dataset achieved test set accuracy: 53.79%

Assignment 3: Fully-Connected Neural Network; Convolutional Neural Network

Status: Completed March 2024
Link to GitHub Repo (Please request access to this private Repo)

Implementation of these classes and functions:

• Linear Fully Connected; ReLU; Linear_ReLU; DropOut layers for image data in tensors
• TwoLayerNet; FullyConnectedNet with arbitrary number of hidden layers
• SGD; SGD w/ momentum; RMSProp; Adam weight updates
• Conv; MaxPool; BatchNorm; SpatialBatchNorm layers for image data with dimensions (num_images, channels, height, width)
• ThreeLayerConvNet; DeepConvNet with arbitrary number of convolutional layers
• Kaiming weights initialization

Key concepts and skills used:

• Define Autograd-like functions with forward and backward calculations
• Use Modules approach to create Fully Connected Networks and Convolutional Neural Networks with arbitrary number of layers (depth)
• Devise Computational Graphs to create Forward Pass and Backward Backpropagation calculations for Linear, ReLU, DropOut, Convolution, MaxPool, Batch Normalization layers
• Implement KaiMing and Xavier weight initiatlizations for Fully Connected linear weights and Convolutional Filters
• Trainable Weights update rules

Trained DeepConvNet (10 Conv + 1 FC layer) model for CIFAR-10 image dataset achieved test set accuracy: 71.66%

Assignment 4: Object Detection

Status: Not Started

Implementation of these classes and functions: TODO

Key concepts and skills used: TODO

Assignment 5: Image Captioning with Recurrent Neural Networks; Transformers

Status: Not Started

Implementation of these classes and functions: TODO

Key concepts and skills used: TODO

Assignment 6: Variational Autoencoder; Generative Adversairal Networks, Network Visualization, Style Transfer

Status: Not Started

Implementation of these classes and functions: TODO

Key concepts and skills used: TODO