How can I debug my model during training?

The model is defined in a model specification file which is a standard Python script.
This script is loaded by the MLTK train command.

Using Visual Studio Code you can single-step debug your model during model training.

  1. If necessary, install the MLTK

  2. Install Visual Studio Code

  3. Open VSCode, and install the Python VScode extension

  4. From VSCode select File on the top-right, then Open Folder and open the folder containing your model’s Python script

  5. From VSCode, open your model Python script and set a breakpoint

  6. Configure the VSCode Python interpreter to point the one with the MLTK installed If you created a virtual environment for the MLTK, then you should select that Python interpreter,
    e.g. ./mltk_pyvenv/Scripts/python.exe

  7. From VSCode, select the Run and Debug tab, then the create a launch.json link,

  8. To .vscode/launch.json, add:

      "version": "0.2.0",
      "configurations": [
          "name": "Python: train",
          "type": "python",
          "request": "launch",
          "module": "mltk",
          "args": [

    This defines a new Python launch configuration which effectively runs the command:

    cd <folder containing your model>
    mltk train --test ./

    However, with this, we can set breakpoints and single-step through the code.

  9. Launch the Python: train debug configuration to debug your script


Any of the other mltk commands can be debugged in a similar manner, just update the args in the .vscode/launch.json configuration.


If you to debug the data generator callbacks, be sure to set the debug option first, e.g.:

my_model.datagen = ParallelAudioDataGenerator(
  debug=True, # Set this to true to enable debugging of the generator