LEIP Train

leip.train(options: TrainOptions) → None

Generates a Quantization Guided Training script.

Parameters

options (TrainOptions) – Options that configure the training.

Return type

None

class leip.TrainOptions(*, input_path: Path, output_path: CreatedDirectoryPath = './outputTrain', training_config_file: Path = PosixPath('leip_train_config.json'), training_script_path: Path = PosixPath('leip_train.py'), invoke_training: bool = True, step: TrainSteps = TrainSteps.TRAIN)

Options for quantization guided training.

Parameters

  • input_path (Path) –

  • output_path (CreatedDirectoryPath) –

  • training_config_file (Path) –

  • training_script_path (Path) –

  • invoke_training (bool) –

  • step (TrainSteps) –

Return type

None

Quantization Guided Callback

class leip.core.train.quantization_guided_callback.QGMonitoringCallback(epoch_begin=False, epoch_end=False, batch_begin=False, batch_end=False)

This class implements automatic monitoring of quantization guided regularization of the model weights.

on_batch_begin(batch, logs=None)

Called on the beginning of the training batch with the batch number.

on_batch_end(batch, logs=None)

Called on the end of the training batch with the batch number.

on_epoch_begin(epoch, logs=None)

Called on the beginning of the training epoch with the epoch number.

on_epoch_end(epoch, logs=None)

Called on the end of the training epoch with the epoch number.

class leip.core.train.quantization_guided_callback.QuantizationGuidedCallback(callbacks_dict, epoch_begin=False, epoch_end=False, batch_begin=False, batch_end=False)

Callback class to monitor Quantization Guided Training. Subclass of tf.keras.callbacks.Callback.

This class can be subclassed to provide more fine grained control of the updating of lambda parameters during training.

on_batch_begin(batch, logs=None)

Called on the beginning of the training batch with the batch number.

on_batch_end(batch, logs=None)

Called on the end of the training batch with the batch number.

on_epoch_begin(epoch, logs=None)

Called on the beginning of the training epoch with the epoch number.

on_epoch_end(epoch, logs=None)

Called on the end of the training epoch with the epoch number.

Quantization Guided Config Help

leip.core.train.quantization_guided.config_help.config_to_help_string()

config_help needs to be kept in sync with the config[“help”] dict in ./train_template.py.template

Quantization Guided Constraint

class leip.core.train.constraints.quantization_guided_constraint.QuantizationGuidedConstraint(num_bits=8, tolerance=1e-05, quantizer_name='asymmetric')

Quantization Guided Constraint class. Subclass of tf.keras.constraints.Constraint. Constraint to restrict weights to a given range during training.

asymmetric(w)

Implements the asymmetric quantization of the weight tensor w

get_config()

Returns a python dict of the configuration parameters for this constraint instance

get_serialized()

Returns a serialized representation of the constraint as a python dict

quantizer(w)

Calls the appropriate quantizer implementation pointed to by the self.quantizer_name on the weight tensor w passed in

to_json()

Returns a JSON string representation of the constraint

update(num_bits=None, tolerance=None)

Updates the attributes of the instance. New values for either num_bits or tolerance or both can be supplied. This update may be called by the training framework either between epochs or even between batches.

Quantization Guided Regularizer

class leip.core.train.regularizers.quantization_guided_regularizer.QuantizationGuidedCWConv2DRegularizer(*args, **kwargs)

A regularizer class for Conv2D layers (channel_axis=-1) based on QAChannelwiseRegularizer.

class leip.core.train.regularizers.quantization_guided_regularizer.QuantizationGuidedCWDepthwiseConv2DRegularizer(*args, **kwargs)

A regularizer class for DepthwiseConv2D layers (channel_axis=-2) based on QAChannelwiseRegularizer.

class leip.core.train.regularizers.quantization_guided_regularizer.QuantizationGuidedCWRegularizer(*args, **kwargs)

A class for channel-wise regularization.

asymmetric(w)

Channel-wise asymmetric quantizer. Overrides the parent class’s asymmetric quantizer.

symmetric(w)

Channel-wise symmetric quantizer.Overrides the parent class’s symmetric quantizer.

Quantization Guided Utilities

Script containing utilities for attaching regularizer

class leip.core.train.utilities.QuantizationGuidedCWConv2DRegularizer(*args, **kwargs)

A regularizer class for Conv2D layers (channel_axis=-1) based on QAChannelwiseRegularizer.

class leip.core.train.utilities.QuantizationGuidedCWDepthwiseConv2DRegularizer(*args, **kwargs)

A regularizer class for DepthwiseConv2D layers (channel_axis=-2) based on QAChannelwiseRegularizer.

class leip.core.train.utilities.QuantizationGuidedConstraint(num_bits=8, tolerance=1e-05, quantizer_name='asymmetric')

Quantization Guided Constraint class. Subclass of tf.keras.constraints.Constraint. Constraint to restrict weights to a given range during training.

asymmetric(w)

Implements the asymmetric quantization of the weight tensor w

get_config()

Returns a python dict of the configuration parameters for this constraint instance

get_serialized()

Returns a serialized representation of the constraint as a python dict

quantizer(w)

Calls the appropriate quantizer implementation pointed to by the self.quantizer_name on the weight tensor w passed in

to_json()

Returns a JSON string representation of the constraint

update(num_bits=None, tolerance=None)

Updates the attributes of the instance. New values for either num_bits or tolerance or both can be supplied. This update may be called by the training framework either between epochs or even between batches.

class leip.core.train.utilities.TrainOptions(*, input_path: Path, output_path: CreatedDirectoryPath = './outputTrain', training_config_file: Path = PosixPath('leip_train_config.json'), training_script_path: Path = PosixPath('leip_train.py'), invoke_training: bool = True, step: TrainSteps = TrainSteps.TRAIN)

Options for quantization guided training.

Parameters

  • input_path (Path) –

  • output_path (CreatedDirectoryPath) –

  • training_config_file (Path) –

  • training_script_path (Path) –

  • invoke_training (bool) –

  • step (TrainSteps) –

Return type

None

leip.core.train.utilities.apply_quantization(model)

A Helper function that applies the quantization based on the regularizer. The idea is to use the regularizer’s own quantizer to change the weights.

Note that this function assumes that “X_regularizer” is the regularizer for “X”. As long as this convention is followed, custom layers can be accommodated.

T.B.D.: This method may add extra nods to the graph.

leip.core.train.utilities.asymmetric_quantizer_error(num_bits, w)

Computes the quantization error for asymmetric quantization. Note that this quantizer performs zero-nudging.

Parameters

  • num_bits (int) – Number of bits.

  • w (np.ndarray) – Array to measure it quantization

  • error. –

Returns

float

leip.core.train.utilities.attach_regularizers(model_or_path, attch_json_scheme, target_keras_h5_file=None, backend_session_reset=True, use_weights=False, strategy=None, custom_objects={})

Attaches regularizer to model_or_path

Parameters

  • model_or_path (str or tf.keras.Model) –

  • attch_json_scheme (str or dict) –

  • target_keras_h5_file (str or None) – If provided, will save the

  • file. (new TF Keras model to this) –

  • backend_session_reset (bool) – If True will reset the backend session.

  • use_weights (bool) – If True, will use the weights from model_or_path.

  • strategy (None or tf.distributed) – If not None, with deserialize the

  • scope. (model under this strategy's) –

  • custom_objects (dict) – Custom objects of the model.

Returns

The new tf keras model instance.

leip.core.train.utilities.copy_and_invoke(training_script_path, config_path=None, invoke=True)

Copy the template script to the specified location.

If a config_path is specified, edit the config dict in the script to read that in and apply it over the default config values.

If invoke is True, run the script.

leip.core.train.utilities.list_tf_keras_model(model_or_path, return_json=False, **kwargs)

Lists the contents of the TF Keras model. This is a CLI utility.

Parameters

  • model_or_path (str or tf.keras.Model) –

  • return_json (bool) – If True, will return attach json.

  • contains (**kwargs) – custom_objects (dict): Dictionary of custom layer names and their signatures. print_model_json (bool): Set to True to print model json print_variables (bool): Set to True to list model parameters. save_model_json (str): save_regularizers_json (str): attach_regularizer_to_all (str): If not None, and an interpretable json string, will be used for all kernels and biases.

leip.core.train.utilities.strip_model_for_inference(model, custom_objects=None)

Returns the model stripped of any ops not needed for inference.

leip.core.train.utilities.validate_regularizer_attachment_scheme(scheme)

Ensure the value for the regularizer_attachment_scheme is either a string (in which case it is a path to a JSON file), or the scheme itself, which is a list of items, each of which are validated.