本文整理汇总了Python中dragnn.python.network_units.maybe_apply_dropout方法的典型用法代码示例。如果您正苦于以下问题：Python network_units.maybe_apply_dropout方法的具体用法？Python network_units.maybe_apply_dropout怎么用？Python network_units.maybe_apply_dropout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dragnn.python.network_units的用法示例。
在下文中一共展示了network_units.maybe_apply_dropout方法的3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: residual
# 需要导入模块: from dragnn.python import network_units [as 别名]
# 或者: from dragnn.python.network_units import maybe_apply_dropout [as 别名]
def residual(old_input, new_input, dropout_keep_rate, layer_norm):
"""Residual layer combining old_input and new_input.
Computes old_input + dropout(new_input) if layer_norm is None; otherwise:
layer_norm(old_input + dropout(new_input)).
Args:
old_input: old float32 Tensor input to residual layer
new_input: new float32 Tensor input to residual layer
dropout_keep_rate: dropout proportion of units to keep
layer_norm: network_units.LayerNorm to apply to residual output, or None
Returns:
float32 Tensor output of residual layer.
"""
res_sum = old_input + network_units.maybe_apply_dropout(new_input,
dropout_keep_rate,
False)
return layer_norm.normalize(res_sum) if layer_norm else res_sum 示例2: dot_product_attention
# 需要导入模块: from dragnn.python import network_units [as 别名]
# 或者: from dragnn.python.network_units import maybe_apply_dropout [as 别名]
def dot_product_attention(queries, keys, values, dropout_keep_rate, bias=None):
"""Computes dot-product attention.
Args:
queries: a Tensor with shape [batch, heads, seq_len, depth_keys]
keys: a Tensor with shape [batch, heads, seq_len, depth_keys]
values: a Tensor with shape [batch, heads, seq_len, depth_values]
dropout_keep_rate: dropout proportion of units to keep
bias: A bias to add before applying the softmax, or None. This can be used
for masking padding in the batch.
Returns:
A Tensor with shape [batch, heads, seq_len, depth_values].
"""
# [batch, num_heads, seq_len, seq_len]
logits = tf.matmul(queries, keys, transpose_b=True)
if bias is not None:
logits += bias
attn_weights = tf.nn.softmax(logits)
# Dropping out the attention links for each of the heads
attn_weights = network_units.maybe_apply_dropout(attn_weights,
dropout_keep_rate,
False)
return tf.matmul(attn_weights, values) 示例3: mlp
# 需要导入模块: from dragnn.python import network_units [as 别名]
# 或者: from dragnn.python.network_units import maybe_apply_dropout [as 别名]
def mlp(component, input_tensor, dropout_keep_rate, depth):
"""Feed the input through an MLP.
Each layer except the last is followed by a ReLU activation and dropout.
Args:
component: the DRAGNN Component containing parameters for the MLP
input_tensor: the float32 Tensor input to the MLP.
dropout_keep_rate: dropout proportion of units to keep
depth: depth of the MLP.
Returns:
the float32 output Tensor
"""
for i in range(depth):
ff_weights = component.get_variable('ff_weights_%d' % i)
input_tensor = tf.nn.conv2d(input_tensor,
ff_weights,
[1, 1, 1, 1],
padding='SAME')
# Apply ReLU and dropout to all but the last layer
if i < depth - 1:
input_tensor = tf.nn.relu(input_tensor)
input_tensor = network_units.maybe_apply_dropout(input_tensor,
dropout_keep_rate,
False)
return input_tensor