Instructions

Environment

• pip install -r requirements.txt

T2.1

• Model choice: ResNet18 trained on train.part1 with BCE loss
• cd <directory/with/train.py>

Train

• • python train.py --train_dataset_path <path/to/train1/train> --val_dataset_path <path/to/val/val>

Test

• • python test.py --dataset_path <path/to/val/val>
• • Metrics achieved on dataset train.part2: Accuracy = 0.99, Precision = 0.99, Recall = 0.99

T2.2 Baseline

• Model choice: custom autoencoder trained on train.part1 with L1 loss
• cd <directory/with/train.py>

Train

• • python train.py --train_dataset_path <path/to/train1/train> --val_dataset_path <path/to/val/val>

Test

• • python test.py --dataset_path <path/to/val/val>
• • Metrics achieved on dataset train.part2: MSE = 0.236

Possilbe improvements

• SwinUNet transformer for image denoising
• Swin Transformer for image restoration
• Convert mel-spectrograms to audio arrays with known construction parameters, such as sampling rate. Then apply something like Speech denoising WaveNet to remove noise