theanets.regularizers.WeightL1

class theanets.regularizers.WeightL1(pattern=None, weight=0.0)

Decay the weights in a model using an L1 norm penalty.

Notes

This regularizer implements the loss() method to add the following term to the network’s loss function:

\[\frac{1}{|\Omega|} \sum_{i \in \Omega} \|W_i\|_1\]

where \(\Omega\) is a set of “matching” weight parameters, and the L1 norm :math`|cdot|_1` is the sum of the absolute values of the elements in the matrix.

This regularizer tends to encourage the weights in a model to be zero. Nonzero weights are used only when they are able to reduce the other components of the loss (e.g., the squared reconstruction error).

References

[Qiu11]Q. Qiu, Z. Jiang, & R. Chellappa. (ICCV 2011). “Sparse dictionary-based representation and recognition of action attributes.”

Examples

This regularizer can be specified at training or test time by providing the weight_l1 or weight_sparsity keyword arguments:

>>> net = theanets.Regression(...)

To use this regularizer at training time:

>>> net.train(..., weight_sparsity=0.1)

By default all (2-dimensional) weights in the model are penalized. To include only some weights:

>>> net.train(..., weight_sparsity=dict(weight=0.1, pattern='hid[23].w'))

To use this regularizer when running the model forward to generate a prediction:

>>> net.predict(..., weight_sparsity=0.1)

The value associated with the keyword argument can be a scalar—in which case it provides the weight for the regularizer—or a dictionary, in which case it will be passed as keyword arguments directly to the constructor.

__init__(pattern=None, weight=0.0)

Methods

__init__([pattern, weight])
log() Log some diagnostic info about this regularizer.
loss(layers, outputs)
modify_graph(outputs) Modify the outputs of a particular layer in the computation graph.