initializer.rst

Initializers

Initializers provide init values for network parameter blobs. In Caffe, they are called Fillers.

An initializer that does nothing. To initialize with zeros, use a ConstantInitializer.

Set everything to a constant.

value

The value used to initialize a parameter blob. Typically this is set to 0.

An initializer based on [BengioGlorot2010], but does not use the fan-out value. It fills the parameter blob by randomly sampling uniform data from [ − S, S] where the scale $S=\sqrt{3 / F_{\text{in}}}$. Here F_in is the fan-in: the number of input nodes.

Heuristics are used to determine the fan-in: For a ND tensor parameter blob, the product of all the 1 to N-1 dimensions are considered as fan-in, while the last dimension is considered as fan-out.

Initialize each element in the parameter blob as independent and identically distributed Gaussian random variables.

mean

Default 0.

std

Default 1.

Initialize the parameter blob to be a random orthogonal matrix (i.e. W^TW = I), times a scalar gain factor. Based on [Saxe2013].

gain

Default 1. Use $\sqrt{2}$ for layers with ReLU activations.