[BUG] Implementation errors on `TransformerModel`.
#672 aberto em 6 de dez. de 2021
Métricas do repositório
- Stars
- (6.832 estrelas)
- Métricas de merge de PR
- (Mesclagem média 143d 16h) (16 fundiu PRs em 30d)
Description
I have looked into [Transformer](https://github.com/unit8co/darts/blame/master/darts/models/forecasting/transformer_model.py) and I have found some errors.
Frist,
In line 167, 170,
src = self.encoder(src) * math.sqrt(self.input_size)
tgt = self.encoder(tgt) * math.sqrt(self.input_size)
I don't think we have to multiply math.sqrt(self.input_size) to inputs (src or tgt).
Because torch.nn.MultiheadAttention take cares this normalization.
Second,
In line 173 - 174,
x = self.transformer(src=src,
tgt=tgt)
There is no tgt_mask for this prediction. In order to use teacher forcing at training stage, user must feed tgt_mask to forward function (specifically square_subsequent_mask defined below). Otherwise decoder inputs before time t can see future decoder inputs (e.g, t+1, t+2, ...) which doesn't exist at inference stage.
[docs] @staticmethod
def generate_square_subsequent_mask(sz: int) -> Tensor:
r"""Generate a square mask for the sequence. The masked positions are filled with float('-inf').
Unmasked positions are filled with float(0.0).
"""
return torch.triu(torch.full((sz, sz), float('-inf')), diagonal=1)
I'm not sure these things are errors. But, in my opinion, it seems this is not correct.
Thank you!