From 7e8e54e84c63171e748bbf09516fd517e6821ace Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Sat, 20 Mar 2021 18:09:06 +0100
Subject: Inital commit for refactoring to lightning

---
 text_recognizer/networks/ctc.py | 58 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 58 insertions(+)
 create mode 100644 text_recognizer/networks/ctc.py

(limited to 'text_recognizer/networks/ctc.py')

diff --git a/text_recognizer/networks/ctc.py b/text_recognizer/networks/ctc.py
new file mode 100644
index 0000000..af9b700
--- /dev/null
+++ b/text_recognizer/networks/ctc.py
@@ -0,0 +1,58 @@
+"""Decodes the CTC output."""
+from typing import Callable, List, Optional, Tuple
+
+from einops import rearrange
+import torch
+from torch import Tensor
+
+from text_recognizer.datasets.util import EmnistMapper
+
+
+def greedy_decoder(
+    predictions: Tensor,
+    targets: Optional[Tensor] = None,
+    target_lengths: Optional[Tensor] = None,
+    character_mapper: Optional[Callable] = None,
+    blank_label: int = 79,
+    collapse_repeated: bool = True,
+) -> Tuple[List[str], List[str]]:
+    """Greedy CTC decoder.
+
+    Args:
+        predictions (Tensor): Tenor of network predictions, shape [time, batch, classes].
+        targets (Optional[Tensor]): Target tensor, shape is [batch, targets]. Defaults to None.
+        target_lengths (Optional[Tensor]): Length of each target tensor. Defaults to None.
+        character_mapper (Optional[Callable]): A emnist/character mapper for mapping integers to characters.  Defaults
+            to None.
+        blank_label (int): The blank character to be ignored. Defaults to 80.
+        collapse_repeated (bool): Collapase consecutive predictions of the same character. Defaults to True.
+
+    Returns:
+        Tuple[List[str], List[str]]: Tuple of decoded predictions and decoded targets.
+
+    """
+
+    if character_mapper is None:
+        character_mapper = EmnistMapper(pad_token="_")  # noqa: S106
+
+    predictions = rearrange(torch.argmax(predictions, dim=2), "t b -> b t")
+    decoded_predictions = []
+    decoded_targets = []
+    for i, prediction in enumerate(predictions):
+        decoded_prediction = []
+        decoded_target = []
+        if targets is not None and target_lengths is not None:
+            for target_index in targets[i][: target_lengths[i]]:
+                if target_index == blank_label:
+                    continue
+                decoded_target.append(character_mapper(int(target_index)))
+            decoded_targets.append(decoded_target)
+        for j, index in enumerate(prediction):
+            if index != blank_label:
+                if collapse_repeated and j != 0 and index == prediction[j - 1]:
+                    continue
+                decoded_prediction.append(index.item())
+        decoded_predictions.append(
+            [character_mapper(int(pred_index)) for pred_index in decoded_prediction]
+        )
+    return decoded_predictions, decoded_targets
-- 
cgit v1.2.3-70-g09d2