Mean Teacher: Semi-Supervised Learning

Project Description

UCL Module Project: Applied Deep Learning (COMP0197)

Date: April 2023

Examining the performance of the semi-supervised learning algorithm, Mean Teacher, for image semantic segmentation task on the Oxford Pet III dataset, through two experiments.

• Experiment 1: A model was trained on 25% labeled and 75% unlabelled data using Mean Teacher, with an aim to study (i) how its performance improves with more (epochs) training and (ii) how its performance compares to supervised only training models.

• Experiment 2: Two additional models were trained on 5%/95% and 10%/90% labeled/unlabelled splits, using Mean Teacher, with the objective to understand how reducing the number of labeled data in Mean Teacher models, affects their performance. Comparative performance of all the models trained is shown below:

Setup

Setting up a virtual environment

First, clone the repository:

git clone https://github.com/ezermoysis1/mean-teacher-ssl

Change your directory to where you cloned the files:

cd mean-teacher-ssl

Create a virtual environment with Python 3.6 or above:

virtualenv venv --python=python3.7 (or python3.7 -m venv venv or conda create -n multiqa python=3.7)

Activate the virtual environment. You will need to activate the venv environment in each terminal in which you want to use the project.

source venv/bin/activate (or source venv/bin/activate.csh or conda activate multiqa)

Install the required dependencies:

pip install -r requirements.txt

Use the code

Data

The dataset used in this paper is The Oxford Pet III dataset. It contains a collection of 7,365 images of cats and dogs and their labels. There are image-level labels, specifying what animal (i.e. dog or cat) and breed is shown, and pixel-level labels, providing a semantic segmentation, referred to as "ground truth mask", of the input images. Click here to download the data.

Train

To train a model with with L = 25% labelled and (1-L) = 75% unlabelled data (referred to as 'M-25' in the report) use the following script. This also trains two benchmark supervised models only, with L% labelled data (referred to as 'M25L' or 'M-25L' in the report, and 100% data respectively ('MU' or M-100').

python main_pt.py 0.25

To train models with different labelled/unlabelled data splits, change the float after the .py. For example to train a L=35% and the two benchmark models, use the following

python main_pt.py 0.35

Evaluation

To evaluate all of the trained models run the following:

python main_pt.py evaluate

Authors

• @ezermoysis1
• @fclarke1

Documentation

Please read the full report of the project here