Development Workflow

Development workflow

1. We start with MIMIC data that's been converted to the CLIF standard: https://physionet.org/content/mimic-iv-ext-clif. We first collate and tokenize it with the cocoa package.

   cocoa pipeline \
       --raw-data-home /path/to/raw \
       --processed-data-home ./processed/dev \
       --verbose
   

2. Next we train a model on this data (with hyperparameter tuning):

   cotorra tune \
       --processed-data-home ./processed/dev \
       --output-home ./output/dev/ \
       --verbose
   

3. You can get generative predictions with:

   cotorra generative-score \
       --processed-data-home ./processed/dev \
       --model-home ./output/dev/mdl-cotorra-tuning \
       --verbose
   

4. You can get representations of the initial parts of the sequences and rep-based predictions with:

   cotorra extract \
       --processed-data-home ./processed/dev \
       --model-home ./output/dev/mdl-cotorra-tuning
   
   cotorra rep-based-score \
       --processed-data-home ./processed/dev \
       --model-home ./output/dev/mdl-cotorra-tuning \
       --verbose
   

[!TIP] For this example, we used a small fraction of the whole dataset, allowing commands to complete in a timely manner. For serious use cases, consider using a terminal multiplexer like tmux or screen so that commands will continue to run if your connection is interrupted.