h56cho
October 16, 2020, 2:53pm
1
Hello,
self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
self.first_dropout = Identity()
if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
self.first_dropout = nn.Dropout(config.summary_first_dropout)
self.last_dropout = Identity()
if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
self.last_dropout = nn.Dropout(config.summary_last_dropout)
def forward(
self, hidden_states: torch.FloatTensor, cls_index: Optional[torch.LongTensor] = None
) -> torch.FloatTensor:
"""
Compute a single vector summary of a sequence hidden states.
Args:
hidden_states (:obj:`torch.FloatTensor` of shape :obj:`[batch_size, seq_len, hidden_size]`):
The hidden states of the last layer.
cls_index (:obj:`torch.LongTensor` of shape :obj:`[batch_size]` or :obj:`[batch_size, ...]` where ... are optional leading dimensions of :obj:`hidden_states`, `optional`):
Used if :obj:`summary_type == "cls_index"` and takes the last token of the sequence as classification
So when cls_index is not specified as the argument in SequenceSummary() statement, is the last token of the sequence used for the classification task? The entire sentence for the description is somewhat awkward…
Thanks,