Data Generation Configuration

class model_compression_toolkit.data_generation.DataGenerationConfig(n_iter, optimizer, data_gen_batch_size, initial_lr, output_loss_multiplier, image_granularity=ImageGranularity.AllImages, scheduler_type=None, bn_alignment_loss_type=None, output_loss_type=None, data_init_type=None, layer_weighting_type=None, image_pipeline_type=None, image_normalization_type=None, extra_pixels=0, bn_layer_types=[], last_layer_types=[], image_clipping=True)

Configuration class for data generation.

Initialize the DataGenerationConfig.

Parameters:

  • n_iter (int) – Number of iterations for data generation.

  • optimizer (Any) – The optimizer used for data generation.

  • data_gen_batch_size (int) – Batch size for data generation.

  • initial_lr (float) – Initial learning rate for the optimizer.

  • output_loss_multiplier (float) – Multiplier for the output loss.

  • image_granularity (ImageGranularity) – Granularity of image data generation. Defaults to ImageGranularity.AllImages.

  • scheduler_type (SchedulerType) – Type of scheduler for the optimizer. Defaults to None.

  • bn_alignment_loss_type (BatchNormAlignemntLossType) – Type of BatchNorm alignment loss. Defaults to None.

  • output_loss_type (OutputLossType) – Type of output loss. Defaults to None.

  • data_init_type (DataInitType) – Type of data initialization. Defaults to None.

  • layer_weighting_type (BNLayerWeightingType) – Type of layer weighting. Defaults to None.

  • image_pipeline_type (ImagePipelineType) – Type of image pipeline. Defaults to None.

  • image_normalization_type (ImageNormalizationType) – Type of image normalization. Defaults to None.

  • extra_pixels (Union[int, Tuple[int, int]]) – Extra pixels to add to the input image size. Defaults to 0.

  • bn_layer_types (List) – List of BatchNorm layer types. Defaults to [].

  • last_layer_types (List) – List of layer types. Defaults to [].

  • image_clipping (bool) – Flag to enable image clipping. Defaults to True.

ImageGranularity

class model_compression_toolkit.data_generation.ImageGranularity(value)

An enum for choosing the image dependence granularity when generating images:

ImageWise

BatchWise

AllImages

SchedulerType

class model_compression_toolkit.data_generation.SchedulerType(value)

An enum for choosing the scheduler type for the optimizer:

REDUCE_ON_PLATEAU - Use the ReduceOnPlateau scheduler.

STEP - Use the Step scheduler.

BatchNormAlignemntLossType

class model_compression_toolkit.data_generation.BatchNormAlignemntLossType(value)

An enum for choosing the BatchNorm alignment loss type:

L2_SQUARE - Use L2 square loss for BatchNorm alignment.

OutputLossType

class model_compression_toolkit.data_generation.OutputLossType(value)

An enum for choosing the output loss type: NONE - No output loss is applied. NEGATIVE_MIN_MAX_DIFF - Use the mean of the negative min-max difference as the output loss. INVERSE_MIN_MAX_DIFF - Use mean of the 1/(min-max) difference as the output loss. REGULARIZED_MIN_MAX_DIFF - Use regularized min-max difference as the output loss.

DataInitType

class model_compression_toolkit.data_generation.DataInitType(value)

An enum for choosing the image dependence granularity when generating images:

Gaussian

Diverse

BNLayerWeightingType

class model_compression_toolkit.data_generation.BNLayerWeightingType(value)

An enum for choosing the layer weighting type:

AVERAGE - Use the same weight per layer.

FIRST_LAYER_MULTIPLIER - Use a multiplier for the first layer, all other layers with the same weight.

GRAD - Use gradient-based layer weighting.

ImagePipelineType

class model_compression_toolkit.data_generation.ImagePipelineType(value)

An enum for choosing the image pipeline type for image manipulation:

SMOOTHING_AND_AUGMENTATION - Apply a smoothing filter, then crop and flip the images.

IDENTITY - Do not apply any manipulation (identity transformation).

ImageNormalizationType

class model_compression_toolkit.data_generation.ImageNormalizationType(value)

An enum for choosing the image normalization type:

TORCHVISION - Normalize the images using torchvision normalization.

KERAS_APPLICATIONS - Normalize the images using keras_applications imagenet normalization.

NO_NORMALIZATION - Do not apply any normalization.