本文整理汇总了Python中tensor2tensor.layers.common_attention.encoder_decoder_attention_loss方法的典型用法代码示例。如果您正苦于以下问题：Python common_attention.encoder_decoder_attention_loss方法的具体用法？Python common_attention.encoder_decoder_attention_loss怎么用？Python common_attention.encoder_decoder_attention_loss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tensor2tensor.layers.common_attention的用法示例。
在下文中一共展示了common_attention.encoder_decoder_attention_loss方法的2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: body
# 需要导入模块: from tensor2tensor.layers import common_attention [as 别名]
# 或者: from tensor2tensor.layers.common_attention import encoder_decoder_attention_loss [as 别名]
def body(self, features):
"""Transformer main model_fn.
Args:
features: Map of features to the model. Should contain the following:
"inputs": Transformer inputs [batch_size, input_length, hidden_dim]
"targets": Target decoder outputs.
[batch_size, decoder_length, hidden_dim]
"target_space_id": A scalar int from data_generators.problem.SpaceID.
Returns:
Final decoder representation. [batch_size, decoder_length, hidden_dim]
"""
hparams = self._hparams
losses = []
if self.has_input:
inputs = features["inputs"]
target_space = features["target_space_id"]
encoder_output, encoder_decoder_attention_bias = self.encode(
inputs, target_space, hparams, features=features, losses=losses)
else:
encoder_output, encoder_decoder_attention_bias = (None, None)
targets = features["targets"]
targets_shape = common_layers.shape_list(targets)
targets = common_layers.flatten4d3d(targets)
decoder_input, decoder_self_attention_bias = transformer_prepare_decoder(
targets, hparams, features=features)
decoder_output = self.decode(
decoder_input,
encoder_output,
encoder_decoder_attention_bias,
decoder_self_attention_bias,
hparams,
nonpadding=features_to_nonpadding(features, "targets"),
losses=losses)
expected_attentions = features.get("expected_attentions")
if expected_attentions is not None:
attention_loss = common_attention.encoder_decoder_attention_loss(
expected_attentions, self.attention_weights,
hparams.expected_attention_loss_type,
hparams.expected_attention_loss_multiplier)
return decoder_output, {"attention_loss": attention_loss}
ret = tf.reshape(decoder_output, targets_shape)
if losses:
return ret, {"extra_loss": tf.add_n(losses)}
else:
return ret 示例2: body
# 需要导入模块: from tensor2tensor.layers import common_attention [as 别名]
# 或者: from tensor2tensor.layers.common_attention import encoder_decoder_attention_loss [as 别名]
def body(self, features):
"""Transformer main model_fn.
Args:
features: Map of features to the model. Should contain the following:
"inputs": Transformer inputs.
[batch_size, input_length, 1, hidden_dim].
"targets": Target decoder outputs.
[batch_size, decoder_length, 1, hidden_dim]
"target_space_id": A scalar int from data_generators.problem.SpaceID.
Returns:
Final decoder representation. [batch_size, decoder_length, hidden_dim]
"""
hparams = self._hparams
losses = []
if self.has_input:
inputs = features["inputs"]
target_space = features["target_space_id"]
encoder_output, encoder_decoder_attention_bias = self.encode(
inputs, target_space, hparams, features=features, losses=losses)
else:
encoder_output, encoder_decoder_attention_bias = (None, None)
targets = features["targets"]
targets_shape = common_layers.shape_list(targets)
targets = common_layers.flatten4d3d(targets)
decoder_input, decoder_self_attention_bias = transformer_prepare_decoder(
targets, hparams, features=features)
decoder_output = self.decode(
decoder_input,
encoder_output,
encoder_decoder_attention_bias,
decoder_self_attention_bias,
hparams,
nonpadding=features_to_nonpadding(features, "targets"),
losses=losses)
expected_attentions = features.get("expected_attentions")
if expected_attentions is not None:
attention_loss = common_attention.encoder_decoder_attention_loss(
expected_attentions, self.attention_weights,
hparams.expected_attention_loss_type,
hparams.expected_attention_loss_multiplier)
return decoder_output, {"attention_loss": attention_loss}
ret = tf.reshape(decoder_output, targets_shape)
if losses:
return ret, {"extra_loss": tf.add_n(losses)}
else:
return ret