Remove zero and one shortuct initializers.

This is for consistency: now *all* initializers are "created" using
call-syntax, as in `initializer()`. Previously this was the case for all
initializers except `zero` and `one`, which is confusing.

Author: lucasb-eyer

beacon8/init/Const.py

@@ -5,6 +5,3 @@ def const(value):
     def init(shape, fan):
         return _np.full(shape, value)
     return init
-
-zero = const(0)
-one = const(1)

beacon8/init/__init__.py

@@ -1,4 +1,4 @@
-from .Const import const, zero, one
+from .Const import const
 from .Xavier import xavier, xavierN, xavierSigm, xavierSigmN
 from .PReLU import prelu, preluN
 from .Normal import normal

beacon8/layers/BatchNormalization.py

@@ -1,5 +1,5 @@
 from .Module import Module
-from beacon8.init import zero, one
+from beacon8.init import const
 from beacon8.utils import create_param, create_param_and_grad
 
 import numpy as _np
@@ -11,15 +11,15 @@ class BatchNormalization(Module):
     def __init__(self, n_features, eps=None):
         Module.__init__(self)
 
-        self.weight, self.grad_weight = create_param_and_grad(n_features, one, 'W_BN')
-        self.bias, self.grad_bias = create_param_and_grad(n_features, zero, 'b_BN')
+        self.weight, self.grad_weight = create_param_and_grad(n_features, const(1), 'W_BN')
+        self.bias, self.grad_bias = create_param_and_grad(n_features, const(0), 'b_BN')
 
-        self.inference_weight = create_param(n_features, one, 'W_BN_inf')
-        self.inference_bias = create_param(n_features, zero, 'b_BN_inf')
+        self.inference_weight = create_param(n_features, const(1), 'W_BN_inf')
+        self.inference_bias = create_param(n_features, const(0), 'b_BN_inf')
 
         # These are buffers for collecting the minibatch statistics.
-        self.buffer_variance = create_param(n_features, one, 'BN_var')
-        self.buffer_mean = create_param(n_features, zero, 'BN_mean')
+        self.buffer_variance = create_param(n_features, const(1), 'BN_var')
+        self.buffer_mean = create_param(n_features, const(0), 'BN_mean')
         self.buffer_counts = _th.shared(_np.asarray(0, dtype=_th.config.floatX))
 
         self.eps = eps or 1e-5

beacon8/layers/Linear.py

@@ -1,5 +1,5 @@
 from .Module import Module
-from beacon8.init import zero, xavier
+from beacon8.init import const, xavier
 from beacon8.utils import create_param_and_grad
 
 import numpy as _np
@@ -8,7 +8,7 @@
 
 class Linear(Module):
 
-    def __init__(self, nin, nout, with_bias=True, initW=xavier(), initB=zero):
+    def __init__(self, nin, nout, with_bias=True, initW=xavier(), initB=const(0)):
         Module.__init__(self)
 
         self.nin = nin

beacon8/layers/SpatialConvolution.py

@@ -1,13 +1,13 @@
 from .Module import Module
-from beacon8.init import zero, xavier
+from beacon8.init import const, xavier
 from beacon8.utils import create_param_and_grad
 
 import theano as _th
 import numpy as _np
 
 
 class SpatialConvolution(Module):
-    def __init__(self, n_input_plane, n_output_plane, k_w, k_h, d_w=1, d_h=1, with_bias=True, initW=xavier(), initB=zero, border_mode='valid', imshape=None):
+    def __init__(self, n_input_plane, n_output_plane, k_w, k_h, d_w=1, d_h=1, with_bias=True, initW=xavier(), initB=const(0), border_mode='valid', imshape=None):
         Module.__init__(self)
         self.n_input_plane = n_input_plane
         self.n_output_plane = n_output_plane

beacon8/layers/SpatialConvolutionCUDNN.py

@@ -1,5 +1,5 @@
 from .Module import Module
-from beacon8.init import zero, xavier
+from beacon8.init import const, xavier
 from beacon8.utils import create_param_and_grad
 
 import theano as _th
@@ -8,7 +8,7 @@
 
 
 class SpatialConvolutionCUDNN(Module):
-    def __init__(self, n_input_plane, n_output_plane, k_w, k_h, d_w=1, d_h=1, pad_w=0, pad_h=0, with_bias=True, initW=xavier(), initB=zero):
+    def __init__(self, n_input_plane, n_output_plane, k_w, k_h, d_w=1, d_h=1, pad_w=0, pad_h=0, with_bias=True, initW=xavier(), initB=const(0)):
         Module.__init__(self)
         self.n_input_plane = n_input_plane
         self.n_output_plane = n_output_plane