summaryrefslogtreecommitdiff
path: root/text_recognizer/criterions/label_smoothing_loss.py
diff options
context:
space:
mode:
authorGustaf Rydholm <gustaf.rydholm@gmail.com>2021-07-29 23:59:52 +0200
committerGustaf Rydholm <gustaf.rydholm@gmail.com>2021-07-29 23:59:52 +0200
commit34098ccbbbf6379c0bd29a987440b8479c743746 (patch)
treea8c68e3036503049fc7034c677ec855465f7a8e0 /text_recognizer/criterions/label_smoothing_loss.py
parentc032ffb05a7ed86f8fe5d596f94e8997c558cae8 (diff)
Configs, refactor with attrs, fix attr bug in iam
Diffstat (limited to 'text_recognizer/criterions/label_smoothing_loss.py')
-rw-r--r--text_recognizer/criterions/label_smoothing_loss.py42
1 files changed, 0 insertions, 42 deletions
diff --git a/text_recognizer/criterions/label_smoothing_loss.py b/text_recognizer/criterions/label_smoothing_loss.py
deleted file mode 100644
index 40a7609..0000000
--- a/text_recognizer/criterions/label_smoothing_loss.py
+++ /dev/null
@@ -1,42 +0,0 @@
-"""Implementations of custom loss functions."""
-import torch
-from torch import nn
-from torch import Tensor
-import torch.nn.functional as F
-
-
-class LabelSmoothingLoss(nn.Module):
- """Label smoothing cross entropy loss."""
-
- def __init__(
- self, label_smoothing: float, vocab_size: int, ignore_index: int = -100
- ) -> None:
- assert 0.0 < label_smoothing <= 1.0
- self.ignore_index = ignore_index
- super().__init__()
-
- smoothing_value = label_smoothing / (vocab_size - 2)
- one_hot = torch.full((vocab_size,), smoothing_value)
- one_hot[self.ignore_index] = 0
- self.register_buffer("one_hot", one_hot.unsqueeze(0))
-
- self.confidence = 1.0 - label_smoothing
-
- def forward(self, output: Tensor, targets: Tensor) -> Tensor:
- """Computes the loss.
-
- Args:
- output (Tensor): Predictions from the network.
- targets (Tensor): Ground truth.
-
- Shapes:
- outpus: Batch size x num classes
- targets: Batch size
-
- Returns:
- Tensor: Label smoothing loss.
- """
- model_prob = self.one_hot.repeat(targets.size(0), 1)
- model_prob.scatter_(1, targets.unsqueeze(1), self.confidence)
- model_prob.masked_fill_((targets == self.ignore_index).unsqueeze(1), 0)
- return F.kl_div(output, model_prob, reduction="sum")