This repository supports the L3DAS22 IEEE ICASSP Grand Challenge and it is aimed at downloading the dataset, pre-processing the sound files and the metadata, training and evaluating the baseline models and validating the final results. We provide easy-to-use instruction to produce the results included in our paper. Moreover, we extensively commented our code for easy customization.
For further information please refer to the challenge website and to the challenge documentation.
Our code is based on Python 3.7.
To install all required dependencies run:
pip install -r requirements.txt
Follow these instructions to properly create and place the kaggle.json file.
It is possible to download the entire dataset through the script download_dataset.py. This script downloads the data, extracts the archives, merges the 2 parts of task1 train360 files and prepares all folders for the preprocessing stage.
To download run this command:
python download_dataset.py --output_path ./DATASETS --unzip True
This script may take long, especially the unzipping stage.
Alternatively, it is possible to manually download the dataset from Kaggle.
The train360 section of task 1 is split in 2 downloadable files. If you manually download the dataset, you should manually merge the content of the 2 folders. You can use the function download_dataset.merge_train360(). Example:
import download_dataset
train360_path = "path_that_contains_both_train360_parts"
download_dataset.merge_train360(train360_path)
The file preprocessing.py provides automated routines that load the raw audio waveforms and their correspondent metadata, apply custom pre-processing functions and save numpy arrays (.pkl files) containing the separate predictors and target matrices.
Run these commands to obtain the matrices needed for our baseline models:
python preprocessing.py --task 1 --input_path DATASETS/Task1 --training_set train100 --num_mics 1
python preprocessing.py --task 2 --input_path DATASETS/Task2 --num_mics 1 --frame_len 100
The two tasks of the challenge require different pre-processing.
For Task1 the function returns 2 numpy arrays contatining:
- Input multichannel audio waveforms (3d noise+speech scenarios) - Shape: [n_data, n_channels, n_samples].
- Output monoaural audio waveforms (clean speech) - Shape [n_data, 1, n_samples].
For Task2 the function returns 2 numpy arrays contatining:
- Input multichannel audio spectra (3d acoustic scenarios): Shape: [n_data, n_channels, n_fft_bins, n_time_frames].
- Output seld matrices containing the class ids of all sounds present in each 100-milliseconds frame alongside with their location coordinates - Shape: [n_data, n_frames, ((n_classes * n_class_overlaps) + (n_classes * n_class_overlaps * n_coordinates))], where n_class_overlaps is the maximum amount of possible simultaneous sounds of the same class (3) and n_coordinates refers to the spatial dimensions (3).
We provide baseline models for both tasks, implemented in PyTorch. For task 1 we use a Beamforming U-Net and for task 2 an augmented variant of the SELDNet architecture. Both models are based on the single-microphone dataset configuration. Moreover, for task 1 we used only Train100 as training set.
To train our baseline models with the default parameters run:
python train_baseline_task1.py
python train_baseline_task2.py
These models will produce the baseline results mentioned in the paper.
GPU is strongly recommended to avoid very long training times.
Alternatively, it is possible to download our pre-trained models with these commands:
python download_baseline_models.py --task 1 --output_path RESULTS/Task1/pretrained
python download_baseline_models.py --task 2 --output_path RESULTS/Task2/pretrained
These models are also available for manual download here.
We also provide a Replicate interactive demo of both baseline models.
Our evaluation metrics for both tasks are included in the metrics.py script. The functions task1_metric and location_sensitive_detection compute the evaluation metrics for task 1 and task 2, respectively. The default arguments reflect the challenge requirements. Please refer to the above-linked challenge paper for additional information about the metrics and how to format the prediction and target vectors.
Example:
import metrics
task1_metric = metrics.task1_metric(prediction_vector, target_vector)
_,_,_,task2_metric = metrics.location_sensitive_detection(prediction_vector, target_vector)
To compute the challenge metrics for our basiline models run:
python evaluate_baseline_task1.py
python evaluate_baseline_task2.py
In case you want to evaluate our pre-trained models, please add
--model_path path/to/model
to the above commands.
The script validate_submission.py can be used to assess the validity of the submission files shape. Instructions about how to format the submission can be found in the L3das website Use these commands to validate your submissions:
python validate_submission.py --task 1 --submission_path path/to/task1_submission_folder --test_path path/to/task1_test_dataset_folder
python validate_submission.py --task 2 --submission_path path/to/task2_submission_folder --test_path path/to/task2_test_dataset_folder
For each task, this script asserts if:
- The number of submitted files is correct
- The naming of the submitted files is correct
- Only the files to be submitted are present in the folder
- The shape of each submission file is as expected
Once you have valid submission folders, please follow the instructions on the link above to proceed with the submission.