Model Specification


The model specification is a standard Python script that defines everything needed to create, train, and evaluate a machine learning model.

This includes:

  • Model version and description

  • Model architecture/layout defined using the Keras API

  • Dataset and any augmentation settings

  • Classification labels (e.g. yes, no)

  • Training batch size, epochs, optimizer, loss metrics

  • Keras training callbacks

  • Tensorflow-Lite Converter settings for quantizing the model

  • Additional parameters to include in the .tflite model file’s metadata

Model Name

The filename of the model specification script is the name given to the model.

So a model specification script with the file path: ~/workspace/ would have a model name: my_model_v1


The model specification script’s name must only contain alphanumeric or underscore characters

Discoverable Models

The list_mltk_models API returns all MltkModel models found on the Model Search Path. To make this API efficient, it only parses the text of the Python scripts found on the model search path for the delimiter @mltk_model (as opposed to loading each model as a Python module).

Thus, to make your model discoverable by this API, add @mltk_model to your model specification script.

MltkModel Class Instance

The core of the model specification script is the MltkModel class instance.
This is a Python class that contains all of the model settings and properties. Additional properties are added to the MltkModel by inheriting additional model “mixins”.

For example, the following might be added to the top of model specification script:

# Define a new MyModel class which inherits the 
# MltkModel and several mixins
# NOTE: The name of the custom class can by anything
# @mltk_model
class MyModel(
    """My Model's class object"""

# Instantiate the MyModel class
my_model = MyModel()

Here we define our model’s class object: MyModel.

At a minimum, this class must inherit MltkModel.

However, to use this model for training or evaluation, it should also inherit others “mixins” such as:

The rest of the model specification script populates the various parameters of the model instance, e.g.:

# General Settings
my_model.version = 1
my_model.description = 'My model is great!'

# Training Basic Settings
my_model.epochs = 100
my_model.batch_size = 64 
my_model.optimizer = 'adam'

# Dataset Settings
my_model.dataset = speech_commands_v2
my_model.class_mode = 'categorical'
my_model.classes = ['up', 'down', 'left', 'right']

Specification Sections

After the model class instantiation, a typical model specification script contains the following sections:

Model Layout

The actual machine learning model is defined using the Keras API. The model is defined in a function that builds a KerasModel. The model specification then sets the model property TrainMixin.build_model_function to reference the function. At model training time, the model building function is invoked and the built KerasModel is trained using Tensorflow.

For example, a model specification script might have:

def my_model_builder(my_model: MyModel):
    keras_model = Sequential(

    keras_model = Sequential()
        kernel_size=(10, 8),

    return keras_model

# Set the model property to reference the model build function
my_model.build_model_function = my_model_builder

Note about hardcoding model layer parameters

While not required, the my_model argument to the building function should be used over hardcoded values, e.g.:

# Good:
# Dynamically determine the number of dense unit based
# on the number of classes specified in the model properties

# Bad:
# Hardcoding dense units
# If the number of classes changes, 
# then training will likely fail

General Settings

This section contains basic information about the model such as its version and description, e.g.:

my_model.version = 1
my_model.description = 'My model is great!'

Basic Training Settings

This section contains basic training settings such as the number of epochs and batch size, e.g.:

my_model.epochs = 100
my_model.batch_size = 64 
my_model.optimizer = 'adam'

Also see:

Training Keras Callback Settings

Tensorflow/Keras supports various callbacks while training.
The MLTK provides properties for some of the more common callbacks, e.g.:

my_model.reduce_lr_on_plateau['monitor'] = 'accuracy'
my_model.reduce_lr_on_plateau['factor'] =  0.25
my_model.reduce_lr_on_plateau['patience'] = 25

Tensorflow-Lite Converter Settings

At the end of training, a quantized .tflite model file is generated. This is the file that is programmed into the embedded device and executed by Tensorflow-Lite Micro.

The .tflite is generated by the Tensorflow-Lite Converter. The settings for the converter are defined in the model specification script using the model property: TrainMixin.tflite_converter.

For example, the model specification script might have:

my_model.tflite_converter['optimizations'] = [tf.lite.Optimize.DEFAULT]
my_model.tflite_converter['supported_ops'] = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
my_model.tflite_converter['inference_input_type'] = tf.int8
my_model.tflite_converter['inference_output_type'] = tf.int8
my_model.tflite_converter['representative_dataset'] = 'generate'

These settings are used at the end of training to generate the .tflite.
See Model Quantization for more details.

Basic Dataset Settings

Basic dataset settings includes information such as the dataset directory, the “class mode”, the class labels, input shape, etc. These settings are strongly dependent on the type of machine learning model that is being trained and/or evaluated.

At a minimum, the DatasetMixin should be inherited by the model specification’s MltkModel class. This mixin is generic and its properties directly map to the parameters required by (the Tensorflow function that actually trains the model)

Alternatively, the MLTK features other dataset mixins that are specific to the dataset type:

  • AudioDatasetMixin - Used for audio datasets (i.e. the samples in the dataset are audio files)

  • ImageDatasetMixin - Used for image datasets (i.e. the samples in the dataset are images)

These are useful as they manage some of the details required to process these types of datasets (whereas the DatasetMixin is generic and all details must be handled manually).

For example, an audio-based model specification script might have:

mltk_model.dataset = speech_commands_v2
mltk_model.class_mode = 'categorical'
mltk_model.classes = ['up', 'down', 'left', 'right']

Alternatively, an image-based model specification script might have:

mltk_model.dataset = rock_paper_scissors_v1
mltk_model.class_mode = 'categorical'
mltk_model.classes = ['rock', 'paper', 'scissors']
mltk_model.input_shape = (96, 96, 1)

Dataset Augmentation Settings

The AudioDatasetMixin and ImageDatasetMixin mixins are useful because they allow for augmenting the dataset during training (which is useful as it effectively increases the size of the dataset which can lead to a more robust model).

These settings are configured in the AudioDatasetMixin.datagen and ImageDatasetMixin.datagen properties.

For example, an audio-based model specification script might have:

mltk_model.datagen = ParallelAudioDataGenerator(

Alternatively, an image-based model specification script might have:

my_model.datagen = ParallelImageDataGenerator(
    contrast_range=(0.50, 1.70),
    noise=['gauss', 'poisson', 's&p'],

Tensorflow Dataset API

Alternatively, you may use the Tensorflow Dataset API. In this case, only the DatasetMixin is required. Note the MLTK provides additional APIs to aid the creation of the Tensorflow datasets, see Tensorflow Dataset Utilities.

The MLTK comes with reference models using the Tensorflow Dataset API:

Evaluation Settings

While an important aspect of the model specification script is configuring the training settings, it can also be used for configuring the model evaluation settings.

Model evaluation is enabled by inheriting one of the following mixins:

Additional Parameters

Additional parameters can also be added to the model specification.
Model Parameters are user-defined and are added to the “metadata” section of the generated .tflite model file.

For example, a model specification script might have:

# Set the volume (in dB) scaler (i.e. amplitude) to apply to the microphone data. If 0 or omitted, no scaler is applied
mltk_model.model_parameters['volume_db'] = 5.0
# This simulates the amount of time in milliseconds an audio loop takes. This helps to provide a better idea of how well the given model will execute on an embedded device
mltk_model.model_parameters['latency_ms'] = 0
# Console logging level, set to 'debug' to enable verbose logging
mltk_model.model_parameters['log_level'] = 'info'


See the reference models for complete model specification scripts: