uncertaintyASR

build docker

run the following command:

$ docker build -t uncertainty_docker .

run docker

$ docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=0 \
    --rm \
    -v <path-to-repo>/src:/root/asr-python/src \
    -v <path-to-repo>/exp:/root/asr-python/exp \
    -v <path-to-repo>/results:/root/asr-python/results \
    -v <path-to-dataset>/TIDIGITS-ASE:/root/asr-python/TIDIGITS-ASE \
    -it uncertainty_docker \
    python3 /root/asr-python/src/recognizer_torch.py 'NN'

Depending on the model use 'NN', 'dropout', 'BNN2', or 'ensemble'

must at least contain the wav files for which we want to create adverarial examples.

run eval

After calculating the adversarial examples, the evaluation on the uncertainty features can be called via:

 docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=0 \
        --rm \
        -v <path-to-repo>/src:/root/asr-python/src \
        -v <path-to-repo>/exp:/root/asr-python/exp \
        -v <path-to-repo>/results:/root/asr-python/results \
        -v <path-to-dataset>/TIDIGITS-ASE:/root/asr-python/TIDIGITS-ASE \
        -it uncertainty_docker \
        python3 /root/asr-python/src/eval.py