Shadows

class todd.models.shadows.BaseShadow

Bases: Module, ABC

__init__(*args, module, device=None, **kwargs)

Parameters:

• module (Module)

• device (Any)

Return type:

None

forward(module)

Parameters:

module (Module)

Return type:

None

property shadow: dict[str, Tensor]

class todd.models.shadows.EMAShadow

Bases: BuildPreHookMixin, BaseShadow

Exponential Moving Average (EMA) Shadow.

This class represents a shadow model that applies exponential moving average to the input data.

Parameters:

ema – The EMA object used for applying exponential moving average.

A copy of the state dict of the given module is stored as the initial shadow:

>>> import torch.nn as nn
>>> module = nn.Module()
>>> module.register_buffer('p', torch.tensor([1., 2., 3.]))
>>> ema = EMAShadow(module=module, ema=EMA())
>>> dict(ema.shadow)
{'p': tensor([1., 2., 3.])}

The shadow updates according to the model:

>>> module.register_buffer('p', torch.tensor([4., 5., 6.]))
>>> ema(module)
>>> dict(ema.shadow)
{'p': tensor([1.0300, 2.0300, 3.0300])}

__init__(*args, ema, **kwargs)

Parameters:

ema (EMA)

Return type:

None

classmethod build_pre_hook(config, registry, item)

Parameters:

• config (Config)

• registry (RegistryMeta)

• item (Item)

Return type:

Config