SELAB-HCMUS MediaEval 2021 Submission

MediaEval 2021: Emotions and Themes in Music

This repo is based on the SELAB-HCMUS submission for MediaEval 2021.

Our ensemble model achieved the 4th rank at the challenge. See results. Our preliminary working note is available in this link.

Requirements

See requirements.txt

pip install -r requirements.txt

Usage

Please prepair the directory containing the preprocessed mel-spectrogram features (See this repo for more details). Run the following command to train the model:

python main.py --model_dir MODEL_DIR --data_dir DATA_DIR --train TRAIN
               --valid VALID --test TEST --name NAME [NAME ...] --size SIZE
               --forget_rate FORGET_RATE --chunk_size CHUNK_SIZE --cut_size
               CUT_SIZE --batch_size BATCH_SIZE --epoch EPOCH

where:

• model_dir: the path to the directory to save and load model
• data_dir: the path to data directory
• train, valid, test: path to .tsv file with music tag labels
• name: list of types of model used for training supported by timm library (for example: ResNet, MobileNet,...)
• size: the size of converted features
• forget_rate: the forget rate described in our working note
• chunk_size: the number of samples used for evaluation
• cut_size: the size of each sample to be cut for evaluation