crumpets.torch.policy module

class crumpets.torch.policy.NoopPolicy

Bases: object

Just a noop Policy. Use it when you don’t want to modify the lr

step(*args, **kwargs)

class crumpets.torch.policy.PolyPolicy(*args: Any, **kwargs: Any)

Bases: torch.optim.lr_scheduler.

A policy that can be described as a polynomial.

Parameters

• optimizer – an optimizer object

• num_epochs – the number of epochs that this policy is defined for. Don’t use it longer than that, because this might cause unexpected behaviour

• power – power value

• last_epoch – The current state of the policy. This can be used to set the initial state of the policy for instance to change the policy during training.

get_lr()

class crumpets.torch.policy.RampPolicy(*args: Any, **kwargs: Any)

Bases: torch.optim.lr_scheduler.

This Policy increases the learning rate step by step

Parameters

• optimizer – an optimizer object

• ramp_epochs – the value where the plateau is reached

• last_epoch – The current state of the policy. This can be used to set the initial state of the policy for instance to change the policy during training.

get_lr()

step(epoch=None, metrics=None)

class crumpets.torch.policy.ReduceLROnPlateau(*args: Any, **kwargs: Any)

Bases: torch.optim.lr_scheduler.

A policy that reduces the learning rate when the training progress reaches a plateau. It inherits from torch.optim.lr_scheduler.ReduceLROnPlateau and because of that shares the same interface

step(epoch=None, metrics=None)

class crumpets.torch.policy.SigmoidPolicy(*args: Any, **kwargs: Any)

Bases: torch.optim.lr_scheduler.

A policy that can be described as a sigmoid. It can be described using the formula base_lr / (1 + math.exp(self.q * x), where x is last_epoch/num_epochs - 1

Parameters

• optimizer – an optimizer object

• num_epochs – the number of epochs that this policy is defined for. Don’t use it longer than that, because this might cause unexpected behaviour

• q – q value to describe the behaviour of the policy.

• last_epoch – The current state of the policy. This can be used to set the initial state of the policy for instance to change the policy during training.

get_lr()