DD2424-project - Transfer Learning

This repo contains a course project in Deep Learning (DD2424) at KTH Royal Institute of Technology. The goal was to explore trainsfer learning on image classification tasks.

Dataset 🐱 🐶

Project is using the Oxford Pets dataset.
Original image width: 600, height: 400.

Setup 🚀

1. Clone repo: git clone git@github.com:simonwanna/DD2424-project.git && cd DD2424-project
2. Create env: conda env create -f environment.yaml && conda activate dd2424
3. Get the data:

mkdir data && cd data
wget https://thor.robots.ox.ac.uk/~vgg/data/pets/images.tar.gz
wget https://thor.robots.ox.ac.uk/~vgg/data/pets/annotations.tar.gz
tar -xvzf images.tar.gz
tar -xvzf annotations.tar.gz
rm images.tar.gz annotations.tar.gz

Train and test 💻

run bash script: bash preprocess.sh, change arguments in file
run training: bash train.sh, change path/arguments in file
run testing: bash test.sh, choose model in file

Results 📊

Best binary classification of cat vs dog.

• val: 0.986
• test: 0.989

Best multiclass classification of 37 breed classes.

• val: 0.916
• test: 0.925