nn.combinators.FFwdStack

class nn.combinators.FFwdStack(*args, **kwargs)[source]

Bases:

Assemble modules into a feed-forward stack, with linear weights in between

FFwdStack accepts any number of modules as positional arguments, along with the required keyword argument weight_init_func.

The weights placed in between each module will map the size_out of one module with the size_in of the following module. Weights are not placed on the input or output of the stack.

Examples

>>> FFwdStack(mod0, mod1, weight_init_func = lambda s: np.random.normal(size = s))

A stack with two modules and one set of linear weights is generated. The weights will have shape (mod0.size_out, mod1.size_in).

Parameters:

  • *mods (Module) – Any number of modules

  • weight_init_func (Callable) – A function that accepts a tuple defining the shape of a matrix, and returns a matrix of that shape to be used as a set of weights

__init__ = <method-wrapper '__init__' of function object>