Closed jramapuram closed 7 years ago
Erm.. Googling these terms seems to produce a lot of noise. Can you link to proper documentation ?
Yea, sorry. Keras has some links to the papers as well as the implementations: https://github.com/fchollet/keras/blob/master/keras/initializations.py
This can be implemented by adding a few functions in Types.hpp
.
We'll need to implement the following:
nn.weight
should be renamed to nn.randomUniform
.nn.randomNormal
for normalnn.constant
similar to af::constantnn.identity
nn.lecunNormal
nn.lecunUniform
nn.glorotNormal
nn.glorotUniform
The links for lecun and glorot can be found here: https://github.com/fchollet/keras/blob/master/keras/initializers.py
Currently Nodes are defined to have spreads:
This needs to be refactored to provide an initializations class. Examples of initializations are :