Run Tensorboard

• The command to run tensorboard is:
  tensorboard --logdir data/output/lightning_logs

Setup containers

1. Create volume: When using the containerized version, first create a volume to hold the data, provide a name for easier reference, e.g. boombox-data:
   docker volume create --name boombox-data
2. Build the container: Execute in the root directory, for Dockerfile to be found:
   docker build . -t boombox
3. Start Container: To start the container with the volume mounted at with gpu-support, run:
   docker run --gpus all -v boombox-data:/BoomboxTransformer/data -t boombox:latest
   Note: To run without a specified volume, use:
   docker run --gpus all -t boombox:latest

Configuration and Hyperparameters

Configuration handling is done with facebook Hydra, the configuration files are hierarchically structured, like so:

├── config.yaml
├── dataset
│   └── noisy_speech.yaml
├── hparams
│   ├── basic.yaml
│   └── lstm.yaml
├── lightning
│   ├── local.yaml
│   └── remote.yaml
└── logging
    └── tensorboard.yaml

• Local vs remote: In the config.yaml you can specify which setup you want to run, if you run locally without gpu, use local, to run with gpu and half-precision training, enter remote
• Number of gpu's: The number of gpu's used by the model is set under lightning/remote.yaml, the default is 1
• Dataset: Urban8k and Librispeech are downloaded and meta-data is created per default, if you have downloaded them already you can set download: False and create_meta: False under dataset/noisy_speech.yaml to save time