TimeAgnosticNet#

class deepinv.models.TimeAgnosticNet(backbone_net: Module)[source]#

Bases: Reconstructor, TimeMixin

Time-agnostic network wrapper.

Adapts a static image reconstruction network to process time-varying inputs. The image reconstruction network then processes the data independently frame-by-frame.

Flattens time dimension into batch dimension at input, and unflattens at output.

Example:

>>> from deepinv.models import UNet, TimeAgnosticNet
>>> model = UNet(scales=2)
>>> model = TimeAgnosticNet(model)
>>> y = rand(1, 1, 4, 8, 8) # B,C,T,H,W
>>> x_net = model(y, None)
>>> x_net.shape == y.shape
True

Parameters:

backbone_net (torch.nn.Module) – Base network which can only take static inputs (B,C,H,W)
device (torch.device) – cpu or gpu.

forward(y: Tensor, physics: Physics, **kwargs)[source]#

Reconstructs a signal estimate from measurements y

Parameters:

y (Tensor) – measurements [B,C,T,H,W]
physics (deepinv.physics.Physics) – forward operator acting on dynamic inputs