cntk.initializer module

Initializers of parameters.

bilinear(kernel_width, kernel_height)

initializer

Parameters:

• kernel_width (int) – kernel width
• kernel_height (int) – kernel height

Returns:

initializer for Parameter useful for deconvolution layer

glorot_normal(scale=1, output_rank=2147483647, filter_rank=2147483647, seed=None)

initializer

Parameters:

• scale (float) – scale
• output_rank (int) – output rank
• filter_rank (int) – filter rank
• seed (int) – random seed

Returns:

initializer for Parameter initialized to Gaussian distribution with mean 0 and standard deviation $$scale*sqrt(2.0/(fanIn+fanOut))$$

glorot_uniform(scale=1, output_rank=2147483647, filter_rank=2147483647, seed=None)

Glorot initializer

Parameters:

• scale (float) – scale
• output_rank (int) – output rank
• filter_rank (int) – filter rank
• seed (int) – random seed

Returns:

initializer for Parameter initialized to uniform distribution between $$scale*sqrt(6.0/(fanIn+fanOut))*[-1,1]$$

he_normal(scale=1, output_rank=2147483647, filter_rank=2147483647, seed=None)

initializer

Parameters:

• scale (float) – scale
• output_rank (int) – output rank
• filter_rank (int) – filter rank
• seed (int) – random seed

Returns:

initializer for Parameter initialized to Gaussian distribution with mean 0 and standard deviation $$scale*sqrt(2.0/fanIn)$$

he_uniform(scale=1, output_rank=2147483647, filter_rank=2147483647, seed=None)

initializer

Parameters:

• scale (float) – scale
• output_rank (int) – output rank
• filter_rank (int) – filter rank
• seed (int) – random seed

Returns:

initializer for Parameter initialized to uniform distribution between $$scale*sqrt(6.0/fanIn)*[-1,1]$$

initializer_with_rank(initializer, output_rank=None, filter_rank=None)

override output_rank and filter_rank specification in a random initializer constructed without an explicit output_rank and filter_rank specification

Parameters:

• initializer – initializer whose output_rank and filter_rank parameters are to be overriden
• output_rank (int) – new output rank value
• filter_rank (int) – new filter rank value

Returns:

new initializer for Parameter with specified output_rank and filter_rank

normal(scale, output_rank=2147483647, filter_rank=2147483647, seed=None)

Normal initializer

Parameters:

• scale (float) – scale
• output_rank (int) – output rank
• filter_rank (int) – filter rank
• seed (int) – random seed

Returns:

initializer for Parameter initialized to normal distribution with mean 0 and standard deviation scale.

truncated_normal(stdev, seed=None)

Truncated normal initializer. The resulting values are drawn from a truncated normal distribution and are always within two standard deviations from the mean (which is 0).

Parameters:

• stdev (float) – standard deviation
• seed (int) – random seed

Returns:

initializer for Parameter initialized to truncated Gaussian distribution between 2*stdev[-1, 1]

uniform(scale, seed=None)

Uniform initializer

Parameters:

• scale (float) – scale
• seed (int) – random seed

Returns:

initializer for Parameter initialized to uniform distribution between scale*[-1.0, 1.0]. Note: this maps to the “uniform1” distribution in BrainScript.

xavier(scale=1, output_rank=2147483647, filter_rank=2147483647, seed=None)

Xavier initializer

Parameters:

• scale (float) – scale
• output_rank (int) – output rank
• filter_rank (int) – filter rank
• seed (int) – random seed

Returns:

initializer for Parameter initialized to Gaussian distribution with mean 0 and standard deviation $$scale*sqrt(3.0/fanIn)$$