Model Training

Training a model is essentially a continuous loop of making a random model and then comparing it's effectiveness to the current best.

After all testing and training data is prepared, training can be initiated with:

    python3 -m train.model -M $model -y $tagged_clips -n $NO_MATCH_clips

This will continue until target_accuracy (from config.yml) is met.

    python3 -m train.model -M clap -y _workspace/tagged/clap -n _workspace/tagged/no-match
    INFO:Training iteration 1
    INFO:Overall accuracy[1] is 50.0
    INFO:Accuracy increased; keeping new model
    INFO:Training iteration 2
    INFO:Overall accuracy[2] is 50.0
    INFO:Accuracy worse than #1; discarding new
    INFO:Training iteration 3
    [...]
    INFO:Training iteration 11
    INFO:Overall accuracy[11] is 84.84848484848484
    INFO:Accuracy worse than #9; discarding new
    INFO:Training iteration 12
    INFO:Overall accuracy[12] is 86.36363636363637
    INFO:Accuracy increased; keeping new model
    INFO:TRAINING COMPLETE :: Final Accuracy: 86.36363636363637

The final products of this training process are model.pth and model.wav. These two files can be copied into another workspace and then used for content inspection (detection).