neural_compressor.objective

The objectives supported by neural_compressor, which is driven by accuracy.

To support new objective, developers just need implement a new subclass in this file.

Module Contents

Classes

Objective	The base class for precise benchmark supported by neural_compressor.
Accuracy	Configuration Accuracy class. Evaluate the accuracy.
Performance	Configuration Performance class. Evaluate the inference time.
Footprint	Configuration Footprint class.
ModelSize	Configuration ModelSize class. Evaluate the model size.
MultiObjective	The base class for multiple benchmarks supported by neural_compressor.

Functions

objective_registry(cls)	The class decorator used to register all Objective subclasses.
objective_custom_registry(name, obj_cls)	Register a customized objective.

neural_compressor.objective.objective_registry(cls)

The class decorator used to register all Objective subclasses.

Parameters:

cls (object) – The class of register.

Returns:

The class of register.

Return type:

cls (object)

neural_compressor.objective.objective_custom_registry(name, obj_cls)

Register a customized objective.

Example:

from eural_compressor.objective import objective_custom_registry

objective_cls = type(user_objective)
name = user_objective.__class__.__name__
objective_cfg = name if deep_get(self.conf.usr_cfg, "tuning.objective") else [name]
deep_set(self.conf.usr_cfg, user_obj_cfg, objective_cfg)
self.conf.usr_cfg = DotDict(self.conf.usr_cfg)
objective_custom_registry(name, objective_cls)

class neural_compressor.objective.Objective

The base class for precise benchmark supported by neural_compressor.

With built-in objectives, users can compress models with different objectives easily. In special cases, users can also register their own objective classes.

class neural_compressor.objective.Accuracy

Configuration Accuracy class. Evaluate the accuracy.

class neural_compressor.objective.Performance

Configuration Performance class. Evaluate the inference time.

class neural_compressor.objective.Footprint

Configuration Footprint class.

Evaluate the peak size of memory blocks during inference.

class neural_compressor.objective.ModelSize

Configuration ModelSize class. Evaluate the model size.

class neural_compressor.objective.MultiObjective(objectives, accuracy_criterion, metric_criterion=[True], metric_weight=None, obj_criterion=None, obj_weight=None, is_measure=False)

The base class for multiple benchmarks supported by neural_compressor.

Example:

from neural_compressor.objective import MultiObjective
obj = MultiObjective(
    objectives=['accuracy', 'modelsize', 'performance'],
    accuracy_criterion={'relative': 0.1},
    obj_criterion=[True, False, False],
    obj_weight=[0.7, 0.2, 0.1])