Style and API

TensorModules

tltorch.TensorModule is a base class we provide for PyTorch modules parametrized by a tensor decomposition.

They must implement get_decomposition()

which returns the decomposition parametrizing the tensor layer.

In the forward pass, the decomposition should be accessed

through tltorch.TensorModule._process_decomposition(),

not directly as module.factors for instance.

This is to make sure all the decomposition hooks are applied first (i.e. tensor dropout).

Initialization

Modules should follow the following API for initialization: for initializating an instance,

use init_from_random, init_from_decompostion and init_from_tensor.

Class method should be used to create an instance and initialize it from existing weights or module (e.g. Class.from_tensor, etc).