diff options
Diffstat (limited to 'text_recognizer/models/transformer.py')
| -rw-r--r-- | text_recognizer/models/transformer.py | 18 |
1 files changed, 9 insertions, 9 deletions
diff --git a/text_recognizer/models/transformer.py b/text_recognizer/models/transformer.py index 75f7523..50bf73d 100644 --- a/text_recognizer/models/transformer.py +++ b/text_recognizer/models/transformer.py @@ -52,16 +52,16 @@ class TransformerLitModel(BaseLitModel): data, targets = batch # Compute the loss. - logits = self.network(data, targets[:-1]) - loss = self.loss_fn(logits, targets[1:]) + logits = self.network(data, targets[:, :-1]) + loss = self.loss_fn(logits, targets[:, 1:]) self.log("val/loss", loss, prog_bar=True) # Get the token prediction. - pred = self(data) - self.val_cer(pred, targets) - self.log("val/cer", self.val_cer, on_step=False, on_epoch=True, prog_bar=True) - self.test_acc(pred, targets) - self.log("val/acc", self.test_acc, on_step=False, on_epoch=True) + # pred = self(data) + # self.val_cer(pred, targets) + # self.log("val/cer", self.val_cer, on_step=False, on_epoch=True, prog_bar=True) + # self.test_acc(pred, targets) + # self.log("val/acc", self.test_acc, on_step=False, on_epoch=True) def test_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None: """Test step.""" @@ -98,8 +98,8 @@ class TransformerLitModel(BaseLitModel): for Sy in range(1, self.max_output_len): context = output[:, :Sy] # (B, Sy) - logits = self.network.decode(z, context) # (B, Sy, C) - tokens = torch.argmax(logits, dim=-1) # (B, Sy) + logits = self.network.decode(z, context) # (B, C, Sy) + tokens = torch.argmax(logits, dim=1) # (B, Sy) output[:, Sy : Sy + 1] = tokens[:, -1:] # Early stopping of prediction loop if token is end or padding token. |