Python version: 3.12.12

To create a virtual environment using python:

# macOS/Linux:
python3.12 -m venv ml-exercise-group19

# Windows:
py -3.12 -m venv ml-exercise-group19

To activate the environment:

# macOS/Linux:
source ml-exercise-group19/bin/activate

# Windows (Command prompt):
ml-exercise-group19\Scripts\activate

To install dependencies:

pip install --upgrade pip
pip install -r requirements.txt

To verify installation:

python test_dependencies.py

Or run one-by-one if it doesn't work:

pip install _

Additionally, if the system has a cuda device (https://developer.nvidia.com/cuda/gpus), you can install the specific pytorch-cuda version (https://pytorch.org/get-started/locally/) and the GPU will be used by default by all deep learning training scripts.

To deactivate the environment:

deactivate

Otherwise, use preferred environment / venv manager and install "requirements.txt".

The CIFAR-10 dataset needs to be downloaded (https://www.cs.toronto.edu/~kriz/cifar.html; CIFAR-10 python version) so that the "cifar-10-batches-py" folder is right inside the "data" folder. The FashionMNIST dataset is loaded from torch.

(1) Traditional Methods

First navigate to the root folder of the project. The file "run_traditional.py" executes all the python scripts and utilizes all feature selection, traditional models and dataset options. The presets are determined in config_traditional.yaml.

python -m src.traditional.run_traditional --config config/config_traditional.yaml

All the results will be saved in the root "results" folder.

(2) Deep Learning

First navigate to src/training. "train.py" uses "config.yml" by default, but using the "--config" argument allows you to select any preset you want. We create presets for every model - dataset - augmentation combination and they can be run with the following commands:

python train.py --config ../../config/sqn_c10_aug.yml
python train.py --config ../../config/sqn_c10_noaug.yml
python train.py --config ../../config/sqn_fmnst_aug.yml
python train.py --config ../../config/sqn_fmnst_noaug.yml
python train.py --config ../../config/rnt_c10_aug.yml
python train.py --config ../../config/rnt_c10_noaug.yml
python train.py --config ../../config/rnt_fmnst_aug.yml
python train.py --config ../../config/rnt_fmnst_noaug.yml

The models will be saved in the root /checkpoints folder.

Note that the ResNet transfer learning models are quite computationally expensive for CPU training and might need less epochs / downsized loaded model to train in a reasonable time frame.

To evaluate the checkpoints, navigate to src/utils. The "eval.py" script takes "--config" and "--checkpoint" as arguments. The checkpoint file has to match the config file with which the model was trained. You can evaluate the model presets with the following commands:

python eval.py --config ../../config/sqn_c10_aug.yml --checkpoint ../../checkpoints/squeezenet_cifar10_aug.pt
python eval.py --config ../../config/sqn_c10_noaug.yml --checkpoint ../../checkpoints/squeezenet_cifar10_noaug.pt
python eval.py --config ../../config/sqn_fmnst_aug.yml --checkpoint ../../checkpoints/squeezenet_fmnist_aug.pt
python eval.py --config ../../config/sqn_fmnst_noaug.yml --checkpoint ../../checkpoints/squeezenet_fmnist_noaug.pt
python eval.py --config ../../config/rnt_c10_aug.yml --checkpoint ../../checkpoints/resnet_cifar10_aug.pt
python eval.py --config ../../config/rnt_c10_noaug.yml --checkpoint ../../checkpoints/resnet_cifar10_noaug.pt
python eval.py --config ../../config/rnt_fmnst_aug.yml --checkpoint ../../checkpoints/resnet_fmnist_aug.pt
python eval.py --config ../../config/rnt_fmnst_noaug.yml --checkpoint ../../checkpoints/resnet_fmnist_noaug.pt

The results will be saved in the root /eval folder.