Keraflow
Deep Learning for Python.
 All Classes Namespaces Functions Pages
keraflow.initializations Namespace Reference

Built-in initialization functions. More...

Functions

def uniform
 Unifrom initialization.
 
def normal
 Normal initialization.
 
def lecun_uniform
 Uniform initialization scaled by the square root of the number of inputs (LeCun 98). More...
 
def glorot_normal
 Gaussian initialization scaled by fan_in + fan_out. More...
 
def glorot_uniform
 Uniform initialization scaled by fan_in + fan_out.
 
def he_normal
 Gaussian initialization scaled by fan_in. More...
 
def he_uniform
 he_uniform initialization.
 
def orthogonal
 Use with square 2D layers (shape[0] == shape[1]). More...
 
def identity
 Use with square 2D layers (shape[0] == shape[1]).
 
def zero
 All zeros initialization.
 
def one
 All ones initialization.
 

Detailed Description

Built-in initialization functions.

Function Documentation

def keraflow.initializations.glorot_normal (   shape,
  name = None 
)

Gaussian initialization scaled by fan_in + fan_out.

  • Reference: Glorot & Bengio, AISTATS 2010
def keraflow.initializations.he_normal (   shape,
  name = None 
)

Gaussian initialization scaled by fan_in.

def keraflow.initializations.lecun_uniform (   shape,
  name = None 
)

Uniform initialization scaled by the square root of the number of inputs (LeCun 98).

def keraflow.initializations.orthogonal (   shape,
  scale = 1.1,
  name = None 
)

Use with square 2D layers (shape[0] == shape[1]).