arnav-gudibande/project-t

CS 189 Project T

This is our submission for CS 189 Project T

Project Members: Charlie Snell, Arnav Gudibande, Divi Schmidt

Installation

You can choose to run this locally or on an instance with a GPU (not required). Once you have an environment setup, download the codebase

• git clone https://github.com/arnav-gudibande/project-t.git
• cd project-t

Install and activate a virtual environment

• python3 -m pip install --user virtualenv
• python3 -m venv env
• source env/bin/activate

Install required dependencies and files

• pip install -r requirements.txt
• python project/finetune/cifar10_download.py

Run Jupyter Notebook

• jupyter notebook project

Navigating Repository

For Students

• project/problem1-3.ipynb - 3 Jupyter notebooks intended to be completed by students
• docs/ - Contains lecture slides, notes, and quiz

For Staff / Maintainers

• project/sol1-3.ipynb - Detailed solution notebooks corresponding to above
• project/finetune/cifar10_models - Source code for CIFAR-10 pre-trained models (credit: https://github.com/huyvnphan/PyTorch_CIFAR10)
• project/data_ft - Contains cached activations for notebook 1

Learning Objectives

• problem1.ipynb - In this notebook, students will discover just how powerful the learned features from pretrained models really are. They will be comparing the performance of 3 different sets of features. The first is the most naive feature set possible: the set of raw pixel values. Then, they will augment the set raw pixel values with a set of features extracted from custom convolutional filters. Lastly, they will use different techniques to compare these features to the features extracted from a state of the art pretrained model. By taking this "feature" based approach, students will be able to leverage concepts from EECS 16AB and earlier parts of 16ML to develop an understanding of the motivation for pre-trained models

• problem2.ipynb - This notebook provides a practical exercise of one of the most common use-cases of pre-trained models -- finetuning. Students will use PyTorch to setup a machine learning workflow and finetune a pre-trained model on the simple task of binary image classification. Then, they will look at a specific case of when pre-trained models really shine -- when there is a lack of available training data. This assignment will not only help build practical skills associated with ML development, but will also build intuition for the situations in which pre-trained models are superior to other alternatives

• problem3.ipynb - This notebook expands on the use-cases for pre-trained models and highlights a variety of situations in which pre-trained models can be used -- fine-tuning on dissimilar datasets, fine-tuning different combinations of layers, etc. The main takeaway for students from this notebook is that there are many use cases for pre-trained models, but there is not a "one-size fits all" solution. Students will also come away with a better understanding for the difference between features in earlier/later layers and other pre-trained models that they can use