diff options
Diffstat (limited to 'text_recognizer')
-rw-r--r-- | text_recognizer/models/vq_transformer.py | 9 |
1 files changed, 5 insertions, 4 deletions
diff --git a/text_recognizer/models/vq_transformer.py b/text_recognizer/models/vq_transformer.py index 71ca2ef..a0d3892 100644 --- a/text_recognizer/models/vq_transformer.py +++ b/text_recognizer/models/vq_transformer.py @@ -1,11 +1,10 @@ """PyTorch Lightning model for base Transformers.""" -from typing import Tuple, Type, Set +from typing import Tuple import attr import torch from torch import Tensor -from text_recognizer.models.metrics import CharacterErrorRate from text_recognizer.models.transformer import TransformerLitModel @@ -13,6 +12,8 @@ from text_recognizer.models.transformer import TransformerLitModel class VqTransformerLitModel(TransformerLitModel): """A PyTorch Lightning model for transformer networks.""" + alpha: float = attr.ib(default=1.0) + def forward(self, data: Tensor) -> Tensor: """Forward pass with the transformer network.""" return self.predict(data) @@ -21,7 +22,7 @@ class VqTransformerLitModel(TransformerLitModel): """Training step.""" data, targets = batch logits, commitment_loss = self.network(data, targets[:-1]) - loss = self.loss_fn(logits, targets[1:]) + commitment_loss + loss = self.loss_fn(logits, targets[1:]) + self.alpha * commitment_loss self.log("train/loss", loss) self.log("train/commitment_loss", commitment_loss) return loss @@ -32,7 +33,7 @@ class VqTransformerLitModel(TransformerLitModel): # Compute the loss. logits, commitment_loss = self.network(data, targets[:-1]) - loss = self.loss_fn(logits, targets[1:]) + commitment_loss + loss = self.loss_fn(logits, targets[1:]) + self.alpha * commitment_loss self.log("val/loss", loss, prog_bar=True) self.log("val/commitment_loss", commitment_loss) |