ls_mlkit.util.observer module

class ls_mlkit.util.observer.Observer(model: Module = None, optimizer: Optimizer = None, scheduler: LambdaLR = None, dataset: Dataset | Dataset = None, target_modules: List[str] = None, no_split_classes: List[str] = None)

Bases: object

function_mapping = {'gradient_norm': <function gradient_norm_fn>, 'gradients': <function gradients_fn>, 'weight_norm': <function weight_norm_fn>, 'weights': <function weights_fn>}

get_gradient_norm(strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None)

get_gradients(strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None)

get_something(name, strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None)

get_something_from_targets(function: Callable)

get_statistics(name, strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None)

get_weight_norm(strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None)

get_weights(strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None)

log_distribution(name, bins: int = 16, section='observer/', strategy: Literal['all', 'block'] = 'all', no_split_classes: List[str] = None, desc='')

log_gradient_norm(section='observer/', strategy: Literal['all', 'block', 'both'] = 'both', no_split_classes: List[str] = None, desc='gradient_norm/')

log_gradients_distribution(bins: int = 16, section='observer/', strategy: Literal['all', 'block', 'both'] = 'both', no_split_classes: List[str] = None, desc='')

log_histograms(bins: int = 16, section='histogram/', prefix='')

log_statistics(strategy: Literal['all', 'block', 'both'] = 'both', no_split_classes: List[str] = None, section='statistics/')

log_weight_norm(section='observer/', strategy: Literal['all', 'block', 'both'] = 'both', desc='', no_split_classes: List[str] = None)

log_weights_distribution(bins: int = 16, section='observer/', strategy: Literal['all', 'block', 'both'] = 'both', no_split_classes: List[str] = None, desc='')

ls_mlkit.util.observer.gradient_norm_fn(module: Module)

Compute the gradient norm of a module

Parameters:

module (Module) – the module to compute the gradient norm

Returns:

the gradient norm of the module

Return type:

float

ls_mlkit.util.observer.gradients_fn(module: Module) → list[Tensor]

Get the gradients of a module

Parameters:

module (Module) – the module to get the gradients

Returns:

the gradients of the module

Return type:

list

ls_mlkit.util.observer.weight_norm_fn(module: Module)

Compute the weight norm of a module

Parameters:

module (Module) – the module to compute the weight norm

Returns:

the weight norm of the module

Return type:

float

ls_mlkit.util.observer.weights_fn(module: Module) → list[Tensor]

Get the weights of a module

Parameters:

module (Module) – the module to get the weights

Returns:

the weights of the module

Return type:

list