Storch Tensors¶
To keep track of the stochastic computation graph, Storchastic returns wrapped torch.Tensor
that are subclasses of
storch.Tensor
. This wrapper contains information that allows Storchastic to analyse the computation graph
during inference to properly estimate gradients. Furthermore, storch.Tensor
contains plate information that allows
for automatic broadcasting with other storch.Tensor
objects with different plate information.
- class storch.tensor.IndependentTensor(tensor: torch.Tensor, parents: [Tensor], plates: [Plate], tensor_name: str, plate_name: str, weight: Optional[storch.Tensor])[source]¶
Bases:
Tensor
Used to denote independencies on a Tensor. This could for example be the minibatch dimension. The first dimension of the input tensor is taken to be independent and added as a batch dimension to the storch system.