本文整理汇总了Python中theano.tensor.shared_randomstreams.RandomStreams.randn方法的典型用法代码示例。如果您正苦于以下问题：Python RandomStreams.randn方法的具体用法？Python RandomStreams.randn怎么用？Python RandomStreams.randn使用的例子？那么恭喜您, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在类theano.tensor.shared_randomstreams.RandomStreams的用法示例。
在下文中一共展示了RandomStreams.randn方法的1个代码示例，这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞，您的评价将有助于系统推荐出更棒的Python代码示例。
示例1: WordEmbeding
# 需要导入模块: from theano.tensor.shared_randomstreams import RandomStreams [as 别名]
# 或者: from theano.tensor.shared_randomstreams.RandomStreams import randn [as 别名]
class WordEmbeding(LayerLab, Block):
def __init__(self, dim, layer_name, input_space=None, use_affine=True,
seed=None, **kwargs):
super(WordEmbeding, self).__init__(**kwargs)
self.dim = dim
self.layer_name = layer_name
self.use_affine = use_affine
def __call__(self, input):
self.get_input_space().validate(input)
return self.upward_pass(input)
def _make_theano_rng(self, seed):
# XXX: Need to handle random Stream management
self.theano_rng = RandomStreams(seed)
@wraps(LayerLab.set_input_space)
def set_input_space(self, space):
if not isinstance(space, IndexSpace):
raise TypeError("WordEmbeding neeeds IndexSpace got: " + str(space))
self.input_space = space
rng = self.mlp.rng
self.output_space = SequenceDataSpace(VectorSpace(dim=self.dim))
self._make_theano_rng(max(rng.randint(2 ** 15), 1))
self._Em = rng.normal(size=(self.input_space.max_labels, self.dim))
self._W = rng.normal(size=(self.input_space.max_labels, self.dim))
self._Em = sharedX(self._Em, name=self.layer_name + '_Em')
self._W = sharedX(self._W, name=self.layer_name + '_W')
self._params = [self._Em, self._W]
if self.use_affine:
self._bias = np.zeros((self.input_space.max_labels,))
self._bias = sharedX(self._bias, name=self.layer_name + '_b')
self._params.append(self._bias)
def upward_pass(self, state_below):
"""
Support Pretrain Layer
"""
self.input_space.validate(state_below)
x = self._Em[state_below] # if input is in
x_shape = x.shape
x = x.reshape((x_shape[1], x_shape[0], x_shape[2]))
return x
def fprop(self, state_below):
if not isinstance(state_below, tuple):
raise TypeError("State_below argument is not a tuple of tensor " +
"variable.\n" + str(state_below))
self.input_space.validate(state_below)
feature, center = state_below
if self.use_affine:
return T.nnet.sigmoid(
T.sum(T.dot(self._Em[feature], self._W[center].T), axis=1) +
self._bias[center])
else:
return self.non_affine(feature, center)
def non_affine_fprop(self, feature, center):
return T.nnet.sigmoid(T.sum(T.dot(self._Em[feature],
self._W[center].T), axis=1))
def _negative_target(self, center):
"""
Return
------
tensor vatiable : sample from word dict
"""
return self.theano_rng.randn(high=self.input_space.max_labels,
size=center.shape)
def nce_loss(self, feature, center):
"""Build the graph for the NCE loss."""
# cross-entropy(logits, labels)
positive_output = self.fprob(feature, center)
negative_target = self._negative_target(center)
negative_output = self.fprob(feature, negative_target)
nce_loss = T.nnet.binary_crossentropy(positive_output,
T.ones_like(center)) + \
T.nnet.binary_crossentropy(negative_output,
T.cast(T.eq(negative_target, positive_output), 'int32'))
# NCE-loss is the sum of the true and noise (sampled words)
# contributions, averaged over the batch.
return nce_loss
def cost(self, feature, target):
return self.nce_loss(feature, target)
@wraps(LayerLab.get_default_cost)
def get_default_cost(self):
return Default()