diff options
Diffstat (limited to 'text_recognizer/networks/transformer/attention.py')
| -rw-r--r-- | text_recognizer/networks/transformer/attention.py | 7 |
1 files changed, 7 insertions, 0 deletions
diff --git a/text_recognizer/networks/transformer/attention.py b/text_recognizer/networks/transformer/attention.py index e1324af..8724691 100644 --- a/text_recognizer/networks/transformer/attention.py +++ b/text_recognizer/networks/transformer/attention.py @@ -58,6 +58,7 @@ class Attention(nn.Module): context_mask: Optional[Tensor], rotary_pos_emb: Optional[Tensor] = None, ) -> Tuple[Tensor, Tensor]: + b, n, _, device = x.shape, x.device q, k, v = self.qkv_fn(x) q, k = ( self._apply_rotary_emb(q, k, rotary_pos_emb) @@ -66,7 +67,13 @@ class Attention(nn.Module): k, ) + input_mask = None if any(x is not None for x in (mask, context_mask)): + q_mask = ( + mask + if mask is not None + else lambda: torch.ones((b, n), device=device).bool() + ) pass # Compute the attention |