DD2424-Project

This is the final version of the code used for the project assignment of course DD2424 given at KTH Royal Institute of Technology.

Setup

For this program to run correctly, the uncompressed Oxford IIIT Pet dataset must be present in the folder named oxford-iiit-pet. Then the following command must be run to generate small, medium and large sized (with respect to the amount of training data) datasets.
(Beware! This will take up ~5 GBs of space in your hard disk)

python generate_datasets.py

Executing the Program

The program can simply be run with the following command:

python image_classification.py

Command-line arguments

Additionally, the following arguments can be given to the program for various effects:

• -t (small|medium|large) := Choose the size of the training dataset (Default large)
• -b := Retrain batch normalization layers (Default True)
• -d := Use data augmentation (Default True)
• -n (1|2|3|4|5) := The number of stages retrained (Default 2)
• -c (2|37) := The number of classes in the dataset (Default 2)
• -e (number) := The number of epochs to train the model for (Default 15)
• --sophisticated_data_augs (none|cutmix|mixup|erase) := Extra option for more sophisticated data augmentations (Default none)
• --only_update_bn_params := Option to train by updating only the batch normalization parameters (Default False)