Error "TypeError: Cannot destructure 'options' as it is undefined" thrown when attempting to create a training call to a private model

[notalexandru]

Description

The trainModel function is throwing a "TypeError: Cannot destructure 'options' as it is undefined" error, even though the options object is passed as an argument to the function.

Code Sample

trainModel = async (filePath, instancePrompt, classPrompt, ...maxTrainSteps) => {
    // Create the training options for the training call
    let inputOptions = this.#createTrainingOptions(filePath, instancePrompt, classPrompt, ...maxTrainSteps);

    // Create a string to store the username of the model
    let callDestination = `${this.#username}/${this.#model}`;

    let options = {
        version: this.#trainerVersion,
        destination: callDestination,
        input: inputOptions,
    };

    // Create a training call to the model
    const response = await this.#replicate.trainings.create(options);

    console.log(response);

} // End trainModel()

The input options arg looks like this:

{
            "instance_prompt": instancePrompt,
            "class_prompt": classPrompt,
            "instance_data": "{file url for training data}",
            "max_train_steps": 500
}

Expected Behavior

The options object should be successfully passed to the trainings.create method without any errors.

Additional Context

The code snippet lacks clear examples or instructions on what parameters should be passed to the trainings.create function. As a result, the API call here is being used as a reference for the required parameters.

[notalexandru]

Just looked through the code that creates the POST request for the training

async function createTraining(model_owner, model_name, version_id, options) {
  const { ...data } = options;

  const training = this.request(`/models/${model_owner}/${model_name}/versions/${version_id}/trainings`, {
    method: 'POST',
    data,
  });

  return training;
}

For some reason it requires three other parameters that are not at all listed in the documentation, those being model_owner, model_name, and version_id. Although this in of itself is not too weird, the fact that it was not at all mentioned in the documentation is incredibly concerning as it seems those parameters are basically required for the trainings.create() function to work.

EDIT: I forgot to add, what is supposed to be included in the input object parameter? As stated previously I was using a command line API call as a reference but not everyone is able to do that. Please update the documentation so that it is in a usable state.

[mattt]

Hi @AlexStan0. Thanks for writing this up. The documentation in the README was out of sync with the actual implementation, but that's now fixed by #83.

For context, training / fine-tuning models is currently in private beta (see this guide for details). DreamBooth predates trainings, and uses a different API that's not supported by this library. Please follow the instructions from this blog post to create custom models with DreamBooth.