Speed up training

Depending on the size and complexity of your original table data, AI model training can take long. You can speed up training times by reducing the maximum training time, the amount of training epochs, decreasing the model size, or increasing the batch size.

The actual training time depends mainly on your data. Because of this, there is no "one-size-fits-all" configuration to reduce training time. Test the configuration of your models and try the suggested configurations in the sections below. Treat the values provided in the examples in each section as a demonstration.

Defaults related to training time

The training speed-related options with their default setting for each model are listed below:

• Max sample size: 100%
• Max training time: 10 min
• Max training epochs: 100
• Model size: Medium
• Batch size: Auto
• Max sequence window (linked tables and Text models only): 100 rows

Use the Speed or Turbo presets

Select the Speed configuration preset for faster AI model training while not compromising on accuracy. The Speed preset applies the following configuration:

• Max sample size: 100%
• Max training time: 10 min
• Max sequence window (linked tables and Text models only): 20 rows

Select the Turbo configuration preset to complete AI model training as quickly as possible at the cost of accuracy. Best for quick sanity checks. The Turbo preset applies the following configuration:

• Max sample size: 100%
• Max training time: 1 min
• Max sequence window (linked tables and Text models only): 4 rows

Decrease training sample size

By default, MOSTLY AI uses all records in a table to train the Generative AI model for that table. Decrease the training sample size if you want to speed up model training.

config = {
    "name": "US census",
    "tables": [
        {
            "data": "https://mostly.ai/docs/datasets/us-census-income.csv",
            "name": "census",
            "modelConfiguration": {
                "maxSampleSize": 1000,  # use only max 1000 samples
            },
        }
    ]
}

Decrease training time

MOSTLY AI sets the default maximum training time to 60 minutes for an AI model. Decrease it to speed up training.

config = {
    "name": "US census",
    "tables": [
        {
            "data": "https://mostly.ai/docs/datasets/us-census-income.csv",
            "name": "census",
            "modelConfiguration": {
                "maxTrainingTime": 1,  # train only for 1 minute
            },
        }
    ]
}

Reduce model size

Model size defines the amount of internal parameters that the AI model uses to learn from your data. A smaller model uses less parameters to analyze and train on your data. You can use three different model sizes.

• Small uses fewer parameters, takes less memory and time, at the cost of accuracy.
• Medium uses optimal parameters and is best for most use cases.
• Large maximizes accuracy with more parameters but requires extra memory and time to complete training.

config = {
    "name": "US census",
    "tables": [
        {
            "data": "https://mostly.ai/docs/datasets/us-census-income.csv",
            "name": "census",
            "modelConfiguration": {
                "modelSize": "S",
            },
        }
    ]
}

Increase batch size

Batch size refers to the number of records used to complete a training pass of the AI model. An epoch completes when all records go through AI model training (depending on the batch size and the records in a table, more than one passes might be necessary).

By default, the batch size is set to Auto and MOSTLY AI determines a batch size that is appropriate for your data. Set a high batch size for a model to speed up model training when accuracy is not a priority.

config = {
    "name": "US census",
    "tables": [
        {
            "data": "https://mostly.ai/docs/datasets/us-census-income.csv",
            "name": "census",
            "modelConfiguration": {
                "batchSize": 1024,
            },
        }
    ]
}

Decrease max sequence window

If you have linked table models (with time-series and events data) in your generator, you can speed up linked model training by decreasing the Max sequence window setting. Bear in mind that depending on your data, this can decrease the accuracy of the model.

For more information, see Configure AI models for time-series and events data.