本文整理汇总了Python中torch.Tensor.masked_select方法的典型用法代码示例。如果您正苦于以下问题：Python Tensor.masked_select方法的具体用法？Python Tensor.masked_select怎么用？Python Tensor.masked_select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torch.Tensor的用法示例。
在下文中一共展示了Tensor.masked_select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: _loss_helper
# 需要导入模块: from torch import Tensor [as 别名]
# 或者: from torch.Tensor import masked_select [as 别名]
def _loss_helper(self, # pylint: disable=inconsistent-return-statements
direction: int,
direction_embeddings: torch.Tensor,
direction_targets: torch.Tensor,
token_embeddings: torch.Tensor) -> Tuple[int, int]:
mask = direction_targets > 0
# we need to subtract 1 to undo the padding id since the softmax
# does not include a padding dimension
# shape (batch_size * timesteps, )
non_masked_targets = direction_targets.masked_select(mask) - 1
# shape (batch_size * timesteps, embedding_dim)
non_masked_embeddings = direction_embeddings.masked_select(
mask.unsqueeze(-1)
).view(-1, self._forward_dim)
# note: need to return average loss across forward and backward
# directions, but total sum loss across all batches.
# Assuming batches include full sentences, forward and backward
# directions have the same number of samples, so sum up loss
# here then divide by 2 just below
if not self._softmax_loss.tie_embeddings or not self._use_character_inputs:
return self._softmax_loss(non_masked_embeddings, non_masked_targets)
else:
# we also need the token embeddings corresponding to the
# the targets
raise NotImplementedError("This requires SampledSoftmaxLoss, which isn't implemented yet.")
# pylint: disable=unreachable
non_masked_token_embeddings = self._get_target_token_embeddings(token_embeddings, mask, direction)
return self._softmax(non_masked_embeddings,
non_masked_targets,
non_masked_token_embeddings)示例2: _get_target_token_embeddings
# 需要导入模块: from torch import Tensor [as 别名]
# 或者: from torch.Tensor import masked_select [as 别名]
def _get_target_token_embeddings(self,
token_embeddings: torch.Tensor,
mask: torch.Tensor,
direction: int) -> torch.Tensor:
# Need to shift the mask in the correct direction
zero_col = token_embeddings.new_zeros(mask.size(0), 1).byte()
if direction == 0:
# forward direction, get token to right
shifted_mask = torch.cat([zero_col, mask[:, 0:-1]], dim=1)
else:
shifted_mask = torch.cat([mask[:, 1:], zero_col], dim=1)
return token_embeddings.masked_select(shifted_mask.unsqueeze(-1)).view(-1, self._forward_dim)