Initializers¶
It is very important to initialize your neural network with correct weights before training. This is not as trivial as it seems as simple initialization like 0, 1, or even the normal distribution usually yield poor results. Most commonly, weights are initialized to be small non-zero values. See this discussion for more info.
-
class
pywick.initializers.
ConstantInitializer
(value, bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
GeneralInitializer
(initializer, bias=False, bias_only=False, **kwargs)[source]¶
-
class
pywick.initializers.
Initializer
[source]¶ Bases:
object
Blank Initializer class from which all other Initializers must inherit
-
class
pywick.initializers.
KaimingNormal
(a=0, mode='fan_in', bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
KaimingUniform
(a=0, mode='fan_in', bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
Normal
(mean=0.0, std=0.02, bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
Orthogonal
(gain=1, bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
Sparse
(sparsity, std=0.01, bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
Uniform
(a=0, b=1, bias=False, bias_only=False, module_filter='*')[source]¶
-
class
pywick.initializers.
XavierNormal
(gain=1, bias=False, bias_only=False, module_filter='*')[source]¶