SUR project 2023/2024 - Audio and Image Classifier

Mária Novákova (xnovak2w), Diana Maxima Držíková (xdrzik01)

The goal of this project is to create classifiers which can detect one specific person in the provided dataset. Two models were created - GMM for audio classification and CNN for image classification.

Predictions for evaluated data:

• for images = models/CNN/trained/predictions_img.txt
• for audio = models/GMM/trained/predictions_audio.txt

Audio Classifier - GMM

We created the Gaussian Mixture Model for audio classification. The model contains two Gaussians - one for the target and non-target. Files are using library ikrlib.py, which was modified to match python3.10.

Three files were created and are handled like this:

Training GMM

python3.10 trainGMM [--cv]

where:

• cv = Stands for the bool parameter which allows cross-validation to be executed. If not called, the default is false.

The models after training are stored in models/GMM/new as gmm_target_model.pkl and gmm_nontarget_model.pkl.

Put SUR_projekt2023-2024 in root folder.

Evaluating GMM

python3.10 evaluateGMM [--input <filepath>] [--output <filepath>] [--target <filepath>] [--non_target <filepath>]

where:

• input = Input files path that are evaluated. Default ../SUR_projekt2023-2024_eval.
• output = Output file path where predictions are stored. Default ../models/GMM/new/predictions_audio.txt.
• target = File with parameters of Gaussian for the target. Default ../models/GMM/trained/gmm_target_model.pkl.
• non_target = File with parameters of Gaussian for non-target. Default ../models/GMM/trained/gmm_nontarget_model.pkl.

Put SUR_projekt2023-2024_eval in root folder.

Library for GMM

libGMM.py is not supposed to be called on it's own. It contains helper functions for training and plotting the results.

Image Classifier - CNN

We created a Convolutional Neural Network for image classification.

Three files were created and are handled like this:

Training CNN

python3.10 trainCNN [--threshold <float>]

where:

• float = Threshold for hard decision. Default 0.5.

The models after training are stored in models/CNN/new as model.pth.

Put SUR_projekt2023-2024 in root folder.

Evaluating CNN

python3.10 evaluateCNN [--input <filepath>] [--output <filepath>] [--model <filepath>] [--threshold <float>]

where:

• input = Input files path that are evaluated. Default ../SUR_projekt2023-2024_eval.
• output = Output file path where predictions are stored. Default ../models/CNN/new/predictions_img.txt.
• model = File with CNN model Default ../models/CNN/new/model.pth.
• threshold = Threshold for hard decision. Default 0.5.

Put SUR_projekt2023-2024_eval in root folder.

Library for CNN

libCNN.py is not supposed to be called on it's own. It contains helper functions for training and plotting the results.