dcase-2020-task1-subtaskB

This repository includes our metadata and code for the submission of IMT Atlantique - BRAIN to the DCASE 2020 challenge, Task 1, subtask B. Our technical report is here

Metadata

For each of the four submitted model :

• metadata of our submissions in yaml files
• Detailed parameter counts, layer-wise (Model_X.csv)
• Detailed loss per category and summary of pruning (log_modelX.txt)

Code

Code is based on pytorch (1.5), sklearn, pandas, numpy, scipy.

1. Resample audio at 18 kHz

Dataset from TAU Urban Acoustic Scenes 2020 3Class can be download at

To resample run,

python resample.py --input_path [path dataset] --output_path [save path]

Note: The output path to the audio dataset has to be specified in create_dataset.py before running the next scripts.

2. Train a model

Launch training

python main_training.py --saving [0 (no), 1 (yes)] --model_type ["ModelA", "ModelB", "ModelC", "ModelD"] --lr [learning rate (float)] --epochs=20 --batch_size=64 --da_mode ['cutmix', 'mixup', 'random_crop'] 

Exemple with ModelB and 5% of the dataset without Data Augmentation (Quick check)

python main_training.py --saving=1 --model_type="ModelB" --lr=1e-3 --epochs=3 --batch_size=32 --frac_data=0.05

Note: Model filename can be found in 'models/' after training with --saving=1

3. Prune and quantify model

Prune model

python pruning_torch.py --model_filename [Model filename] --da_mode ['cutmix', 'mixup', 'random_crop'] 

Quantify model

python quantify_torch.py --model_filename [Model filename]

4. Get statistics and create submission

Get statistics of the model

python model_statistics.py --model_filename [Model filename]

Create submission csv file

python create_submission.py --model_filename [Model filename] --eval_csv_path [Path to the evaluation csv file] --eval_data_path [Path to evaluation audio directory]

Pretrained models

Pretrained models will be uploaded after the challenge deadline.

Team members

• Nicolas Pajusco
• Richard Huang
• Nicolas Farrugia (PI)

Acknowledgments to Carlos Lassance, Ghouthi Boukli Hacene, Vincent Gripon and other members of BRAIN for feedback, comments and informal discussions regarding this submission.