Repo for DCASE 2023 Task 2, anomaly detection.

1. Load conda env with "conda env create -f environment.yml"
2. Activate conda environment with "conda activate dcase2023"
3. If on Mac/Linux install sox with "conda install -c conda-forge sox" (a dependency of torchaudio)

By default, when using source/data_prep.py, full length audio files are segmented into fixed length samples of 1 second (length is configurable). Mel spectrograms are computed for each segmented sample. The segmented audio samples and mel spectrograms are saved.

1. Set up a folder "data/raw" in root directory of this repository
2. Download the ZIP folders for each machine in the Task 2 train set: https://zenodo.org/record/7690157#.ZC7sl3tBzMY
3. Extract the ZIP folders into "data/raw" ("data" folder has to be on same level as "output" and "source" folders)
4. Remove one folder level for each machine such that the directory has the following structure: "raw_path/machine_name/train/...", for example /raw_path/gearbox/train/
5. Run "python source/data_prep.py" to chop samples in "data/raw" into 1 second snippets and save to "data/processed/audio_segments" and spectrograms to "data/processed/spectrograms" The resulting files will have following namings: f"index={index}__segment_id={segment_id}__domain={domain}__label={label}" index: a number indicating the original audio file index (row index in the sorted list of raw audio files) belonged to segment_id is a number from 0 to the number of audio segments produced from one audio file label is "normal" or "anomaly" (only relevant for test data) domain is "source" or "target"

1. Set desired config parameters in "source/config.py"
2. Run "python source/main.py" to start training and testing main.py will train a model for each machine type and produce evaluation results: Two files in each results/machine/ with the anomaly scores and the prediction The results follow the official submission format: "anomaly_score_<machine_type>section<section_index>.csv" with rows: "[filename (string)],[anomaly score (real value)]" "decision_result_<machine_type>section<section_index>.csv" with rows "[filename (string)],[decision result (0: normal, 1: anomaly)]" Every 5 epochs accurracy (which is not really meaningful right now) as well aus auc is tracked with weight and biases as well as in the end all metrics are computed and saved to results/metrics.csv

1. Find reasonable choices for the anomaly score detection threshold (for each machine). This will also be model dependent! For now this is set to a fixed value in the configs DETECTION_TRESHOLD_DICT
2. If the full evaluation pipeline is done, we can start comparing to the baselines/previous years scores and improve the model