Module

Class Module is subclassed from LightningModule of pytorch-lightning.

The differences are list below.

Initialization

Every instance of Module needs a config argument to initialize. It collapses the OptunaSuggest objects in it and becomes an attributes of the instance. So the configuration item can be accessed by self.config.xxx.

Danger

self.config is a FrozenDiot object to prevent the changes to be passed down to other modules (See diot).

on_epoch_end

An empty line is printed on every epoch end to keep the progress bar of the previous epoch run so that we can easily keep track of the metrics of all epochs. If you want to disable this just overwrite on_epoch_end with a pass statement. Or if you want to keep this and do something else in on_epoch_end, you should call super().on_epoch_end() inside your on_epoch_end.

loss_function

Calculate loss using the loss function according to the loss configuration item.

configure_optimizers

You don't have to implement this function if you have optim specified in configuration.

Note

Currently only adam and sgd are supported. Please also specify learning_rate and/or momentum for corresponding optimizers (see configurations).

measure

This method calculates a metric according to num_classes. For num_classes=1 (regression), available metrics are mse, rmse, mae and rmsle. And for classifications, avaiable metrics are accuracy, precision, recall, f1_score, iou, fbeta_score, auroc, average_precision and dice_score.

To use it in your training_step, validation_step or test_step, you can do:

self.measure(output, labels, 'accuracy')

If the metric need extra arguments, you can pass them in as well:

self.measure(output, labels, 'fbeta_score', beta=1.0)

For more details of the metrics, see pytorch-lightning's documentation