StepShift(attr, gamma, step, init=None, target=None, optimizer=None)¶
Trainer extension to shift an optimizer attribute in “steps”.
This extension multiplies the specified attribute of the optimizer in “steps”. The typical use case is to scale the attribute at every
For example, suppose that this extension is invoked at every iteration, then given
k, a multiplier
gammaand an initial value
init, the optimizer attribute is set to
init * gamma ^ (floor(i / k)), where
irepresents the index of the current iteration.
This extension is also called before the training loop starts by default.
- attr (str) – Name of the optimizer attribute to adjust.
- gamma (float) – The multiplier.
- step (int) – The interval for the multiplication, i.e.,
- init (float) – Initial value of the attribute. If it is
None, the extension extracts the attribute at the first call and uses it as the initial value.
- target (float) – Target value of the attribute. If the attribute reaches this value, the shift stops.
- optimizer (Optimizer) – Target optimizer object. If it is None, the main optimizer of the trainer is used.
Invokes the extension.
Implementations should override this operator. This method is called at iterations which the corresponding trigger accepts.
Parameters: trainer (Trainer) – Trainer object that calls this operator.
Finalizes the extension.
This method is called at the end of the training loop.
Initializes up the trainer state.
This method is called before entering the training loop. An extension that modifies the state of
Trainercan override this method to initialize it.
When the trainer has been restored from a snapshot, this method has to recover an appropriate part of the state of the trainer.
ExponentialShiftextension changes the optimizer’s hyperparameter at each invocation. Note that the hyperparameter is not saved to the snapshot; it is the responsibility of the extension to recover the hyperparameter. The
ExponentialShiftextension recovers it in its
initializemethod if it has been loaded from a snapshot, or just setting the initial value otherwise.
Parameters: trainer (Trainer) – Trainer object that runs the training loop.
Serializes the extension state.
It is called when a trainer that owns this extension is serialized. It serializes nothing by default.
Default name of the extension.
It is the name of the class by default. Implementation can override this property, or provide a class attribute to hide it.
trigger= (1, 'iteration')¶