In mlverse/luz: Higher Level 'API' for 'torch'

The ctx object is used in luz to share information between the training loop and callbacks, model methods, and metrics. The table below describes information available in the ctx by default. Other callbacks could potentially modify these attributes or add new ones.

| Attribute | Description | |-------------|-----------------------------------------------------------| | verbose | The value (TRUE or FALSE) attributed to the verbose argument in fit . | | accelerator | Accelerator object used to query the correct device to place models, data, etc. It assumes the value passed to the accelerator parameter in fit. | | model | Initialized nn_module object that will be trained during the fit procedure. | | optimizers | A named list of optimizers used during training. | | data | The currently in-use dataloader. When training it's ctx$train_data, when doing validation its ctx$valid_data. It can also be the prediction dataset when in predict. | | train_data | Dataloader passed to the data argument in fit. Modified to yield data in the selected device. | | valid_data | Dataloader passed to the valid_data argument in fit. Modified to yield data in the selected device. | | min_epochs | Minimum number of epochs the model will be trained for. | | max_epochs | Maximum number of epochs the model will be trained for. | | epoch | Current training epoch. | | iter | Current training iteration. It's reset every epoch and when going from training to validation. | | training | Whether the model is in training or validation mode. See also help("luz_callback_train_valid") | | callbacks | List of callbacks that will be called during the training procedure. It's the union of the list passed to the callbacks parameter and the default callbacks. | | step | Closure that will be used to do one step of the model. It's used for both training and validation. Takes no argument, but can access the ctx object. | | call_callbacks | Call callbacks by name. For example call_callbacks("on_train_begin") will call all callbacks that provide methods for this point. | | batch | Last batch obtained by the dataloader. A batch is a list() with 2 elements, one that is used as input and the other as target. | | input | First element of the last batch obtained by the current dataloader. | | target | Second element of the last batch obtained by the current dataloader. | | pred | Last predictions obtained by ctx$model$forward . Note: can be potentially modified by previously ran callbacks. Also note that this might not be available if you used a custom training step. | | loss_fn | The active loss function that will be minimized during training. | | loss | Last computed loss from the model. Note: this might not be available if you modified the training or validation step. | | opt | Current optimizer, ie. the optimizer that will be used to do the next step to update parameters. | | opt_nm | Current optimizer name. By default it's opt , but can change if your model uses more than one optimizer depending on the set of parameters being optimized. | | metrics | list() with current metric objects that are updated at every on_train_batch_end() or on_valid_batch_end(). See also help("luz_callback_metrics") | | records | list() recording metric values for training and validation for each epoch. See also help("luz_callback_metrics") . Also records profiling metrics. See help("luz_callback_profile") for more information. | | handlers | A named list() of handlers that is passed to rlang::with_handlers() during the training loop and can be used to handle errors or conditions that might be raised by other callbacks. | | epoch_handlers | A named list of handlers that is used with rlang::with_handlers(). Those handlers are used inside the epochs loop, thus you can handle epoch specific conditions, that won't necessarily end training. |

: Context attributes

mlverse/luz documentation built on Sept. 19, 2024, 11:20 p.m.