sparse.nn.trainer.TestMatchingPursuitParameters

class sparse.nn.trainer.TestMatchingPursuitParameters(model: Module, criterion: Module, data_loader: DataLoader, bmp_params_range: Tensor, param_name='param', **kwargs)

Bases: TrainerAutoencoder

Swipe through a range of Softshrink threshold parameters bmp_params_range and show the best one.

The user then can pick the “best” parameter by looking at the PSNR and sparsity plots - the choice is not trivial and depends on the application.

__init__(model: Module, criterion: Module, data_loader: DataLoader, bmp_params_range: Tensor, param_name='param', **kwargs)

Methods

__init__(model, criterion, data_loader, ...)
checkpoint_path([best])	Get the checkpoint path, given the mode.
full_forward_pass([train])	Fixes the model weights, evaluates the epoch score and updates the monitor.
is_unsupervised()	Returns
log_trainer()	Logs the trainer in Visdom text field.
monitor_functions()	Override this method to register Visdom callbacks on each epoch.
open_monitor([offline])	Opens a Visdom monitor.
plot_autoencoder()	Plots AutoEncoder reconstruction.
restore([checkpoint_path, best, strict])	Restores the trainer progress and the model from the path.
save([best])	Saves the trainer and the model parameters to self.checkpoint_path(best).
state_dict()	Returns
train([n_epochs, mutual_info_layers, ...])	User-entry function to train the model for n_epochs.
train_batch(batch)	The core function of a trainer to update the model parameters, given a batch.
train_epoch(epoch)	Trains an epoch.
train_mask([mask_explain_params])	Train mask to see what part of an image is crucial from the network perspective (saliency map).
training_finished()	Training is finished callback.
training_started()	Training is started callback.
update_accuracy([train])	Updates the accuracy of the model.
update_best_score(score)	If score is greater than the self.best_score, save the model.

Attributes

best_score_type
epoch	The current epoch, int.
watch_modules