Project files

All code for the project is stored in the code folder

Strategy 1: Stacked MFCC & Chroma

Files related to strategy one include:

• audio_toolbox/dataset.py: contains the class AudioOTFDataset, used for generating vectorized representations of the audios stored as .wav files on the fly.
• notebooks/dataset_generation/generate_full_dataset.ipynb: the notebook to generate and save the datasets for strategy one.
• notebooks/strategy_one.ipynb: the notebook to preprocess and fit ML models on the dataset.
• notebooks/mlp_strategy_one.ipynb: the notebook to use MLP fit the dataset.

Strategy 2: Manual Feature Engineering

Files related to strategy two include:

• notebooks/dataset_generation/generate_30sec_tabular_dataset.ipynb: the notebook to generate features for audios and save the datasets for strategy 2.1.
• notebooks/tabular_ml_modeling.ipynb.ipynb: the notebook to preprocess and fit ML models on the generated datasets with strategy 2.1.
• audio_toolbox/datasets.py: contains SplitedDataset, which involved audio spliting and feature computation methods.
• notebooks/dataset_generation/generate_split_dataset.ipynb: the notebook to generated splited audios and compute features for both the original audio and the audio segments for strategy 2.2.
• notebooks/split_weight_training_3_sec.ipynb and notebooks/split_weight_training_6_sec.ipynb: the notebook for fitting ML models with strategy 2.2.

Strategy 3: Two-Step VGG fine-tuning

Files related to strategy three include:

• audio_toolbox/datasets.py: contains AudioImageDataset, which is used to generate spectrograms for audios.
• notebooks/dataset_generation/generate_image_dataset.ipynb: the notebook to generate spectrograms for audios and save the datasets for strategy 3.
• notebooks/VGG_fine_tuning.ipynb: the notebook for fine-tuning VGG according to strategy 3.

To run docker container:

• docker-compose up --build
• Go to http://localhost:4000 to see your live site!