controllers module

Full Documentation for hippynn.experiment.controllers module. Click here for a summary page.

Managing optimizer, scheduler, and end of training

class Controller(optimizer, scheduler, batch_size, max_epochs, stopping_key, eval_batch_size=None, fraction_train_eval=0.1, quiet=False)

Bases: object

Class for controlling the training dynamics.

Parameters:

• optimizer – pytorch optimizer (or compatible)

• scheduler – pytorch scheduler (or compatible) Pass None to use no schedulers. May pass a list of schedulers to use them all in sequence.

• scheduler_list – list of schedulers (may be empty).

• stopping_key – str of name for best metric.

• batch_size – batch size for training

• eval_batch_size – batch size during evaluation

• fraction_train_eval – the random fraction of the training set to use during evaluation.

• quiet – If true, suppress printing.

• max_epochs – maximum amount of epochs currently allowed based on training so far.

Note

If a scheduler defines a set_controller method, this will be run on the Controller instance. This allows a scheduler to modify other aspects of training besides the conventional ones found in the optimizer. See RaiseBatchSizeOnPlateau in this module.

load_state_dict(state_dict)

push_epoch(epoch, better_model, metric, _print=<built-in function print>)

state_dict()

property max_epochs

class PatienceController(*args, termination_patience, **kwargs)

Bases: Controller

A subclass of Controller that terminates if training has not improved for a given number of epochs.

Variables:

• patience – How many epochs are allowed without improvement before termination.

• last_best – The eoch number of the last best epoch encountered.

push_epoch(epoch, better_model, metric, _print=<built-in function print>)

property max_epochs

class RaiseBatchSizeOnPlateau(optimizer, max_batch_size, factor=0.5, patience=10, threshold=0.0001, threshold_mode='rel', verbose=None, controller=None)

Bases: ReduceLROnPlateau

Learning rate scheduler compatible with pytorch schedulers.

Note: The “VERBOSE” Parameter has been deprecated and no longer does anything.

This roughly implements the scheme outlined in the following paper:

"Don't Decay the Learning Rate, Increase the Batch Size"
Samuel L. Smith et al., 2018.
arXiv:1711.00489
Published as a conference paper at ICLR 2018.

Until max_batch_size has been reached. After that, the learning rate is decayed.

Note

To use this scheduler, build it, then link it to the container which governs the batch size using set_controller. The default base hippynn Controller will do this automatically.

load_state_dict(state_dict)

Load the scheduler’s state.

set_controller(box)

state_dict()

Return the state of the scheduler as a dict.

It contains an entry for every variable in self.__dict__ which is not the optimizer.

step(metrics)

Perform a step.

property optimizer

is_scheduler_like(thing)

Test if an object can be used as a scheduler. :meta private: