Add greedy decoder

5 files changed, 74 insertions, 55 deletions

diff --git a/README.md b/README.md
index f549cda..6f95001 100644
--- a/README.md
+++ b/README.md
@@ -73,7 +73,7 @@ Ideas of mine that did not work unfortunately:
 - [ ] Evaluation
 - [ ] Wandb artifact fetcher
 - [ ] fix linting
-- [ ] Modularize the decoder
+- [x] Modularize the decoder
 - [ ] Add kv cache
 - [x] Fix stems
 - [x] residual attn
diff --git a/text_recognizer/models/greedy_decoder.py b/text_recognizer/models/greedy_decoder.py
new file mode 100644
index 0000000..9d2f192
--- /dev/null
+++ b/text_recognizer/models/greedy_decoder.py
@@ -0,0 +1,51 @@
+"""Greedy decoder."""
+from typing import Type
+from text_recognizer.data.tokenizer import Tokenizer
+import torch
+from torch import nn, Tensor
+
+
+class GreedyDecoder:
+    def __init__(
+        self,
+        network: Type[nn.Module],
+        tokenizer: Tokenizer,
+        max_output_len: int = 682,
+    ) -> None:
+        self.network = network
+        self.start_index = tokenizer.start_index
+        self.end_index = tokenizer.end_index
+        self.pad_index = tokenizer.pad_index
+        self.max_output_len = max_output_len
+
+    def __call__(self, x: Tensor) -> Tensor:
+        bsz = x.shape[0]
+
+        # Encode image(s) to latent vectors.
+        img_features = self.network.encode(x)
+
+        # Create a placeholder matrix for storing outputs from the network
+        indecies = torch.ones((bsz, self.max_output_len), dtype=torch.long).to(x.device)
+        indecies[:, 0] = self.start_index
+
+        for Sy in range(1, self.max_output_len):
+            tokens = indecies[:, :Sy]  # (B, Sy)
+            logits = self.network.decode(tokens, img_features)  # (B, C, Sy)
+            indecies_ = torch.argmax(logits, dim=1)  # (B, Sy)
+            indecies[:, Sy : Sy + 1] = indecies_[:, -1:]
+
+            # Early stopping of prediction loop if token is end or padding token.
+            if (
+                (indecies[:, Sy - 1] == self.end_index)
+                | (indecies[:, Sy - 1] == self.pad_index)
+            ).all():
+                break
+
+        # Set all tokens after end token to pad token.
+        for Sy in range(1, self.max_output_len):
+            idx = (indecies[:, Sy - 1] == self.end_index) | (
+                indecies[:, Sy - 1] == self.pad_index
+            )
+            indecies[idx, Sy] = self.pad_index
+
+        return indecies
diff --git a/text_recognizer/models/transformer.py b/text_recognizer/models/transformer.py
index 6048901..bbfaac0 100644
--- a/text_recognizer/models/transformer.py
+++ b/text_recognizer/models/transformer.py
@@ -1,6 +1,6 @@
 """Lightning model for base Transformers."""
-from collections.abc import Sequence
-from typing import Optional, Tuple, Type
+from typing import Callable, Optional, Sequence, Tuple, Type
+from text_recognizer.models.greedy_decoder import GreedyDecoder
 
 import torch
 from omegaconf import DictConfig
@@ -20,6 +20,7 @@ class LitTransformer(LitBase):
         loss_fn: Type[nn.Module],
         optimizer_config: DictConfig,
         tokenizer: Tokenizer,
+        decoder: Callable = GreedyDecoder,
         lr_scheduler_config: Optional[DictConfig] = None,
         max_output_len: int = 682,
     ) -> None:
@@ -35,9 +36,10 @@ class LitTransformer(LitBase):
         self.test_cer = CharErrorRate()
         self.val_wer = WordErrorRate()
         self.test_wer = WordErrorRate()
+        self.decoder = decoder
 
     def forward(self, data: Tensor) -> Tensor:
-        """Forward pass with the transformer network."""
+        """Autoregressive forward pass."""
         return self.predict(data)
 
     def teacher_forward(self, data: Tensor, targets: Tensor) -> Tensor:
@@ -59,7 +61,7 @@ class LitTransformer(LitBase):
         logits = self.teacher_forward(data, targets[:, :-1])
         loss = self.loss_fn(logits, targets[:, 1:])
         preds = self.predict(data)
-        pred_text, target_text = self._get_text(preds), self._get_text(targets)
+        pred_text, target_text = self._to_tokens(preds), self._to_tokens(targets)
 
         self.val_acc(preds, targets)
         self.val_cer(pred_text, target_text)
@@ -76,7 +78,7 @@ class LitTransformer(LitBase):
         logits = self.teacher_forward(data, targets[:, :-1])
         loss = self.loss_fn(logits, targets[:, 1:])
         preds = self(data)
-        pred_text, target_text = self._get_text(preds), self._get_text(targets)
+        pred_text, target_text = self._to_tokens(preds), self._to_tokens(targets)
 
         self.test_acc(preds, targets)
         self.test_cer(pred_text, target_text)
@@ -86,55 +88,12 @@ class LitTransformer(LitBase):
         self.log("test/cer", self.test_cer, on_step=False, on_epoch=True, prog_bar=True)
         self.log("test/wer", self.test_wer, on_step=False, on_epoch=True, prog_bar=True)
 
-    def _get_text(
+    def _to_tokens(
         self,
-        xs: Tensor,
-    ) -> Tuple[Sequence[str], Sequence[str]]:
-        return [self.tokenizer.decode(x) for x in xs]
+        indecies: Tensor,
+    ) -> Sequence[str]:
+        return [self.tokenizer.decode(i) for i in indecies]
 
     @torch.no_grad()
     def predict(self, x: Tensor) -> Tensor:
-        """Predicts text in image.
-
-        Args:
-            x (Tensor): Image(s) to extract text from.
-
-        Shapes:
-            - x: :math: `(B, H, W)`
-            - output: :math: `(B, S)`
-
-        Returns:
-            Tensor: A tensor of token indices of the predictions from the model.
-        """
-        start_index = self.tokenizer.start_index
-        end_index = self.tokenizer.end_index
-        pad_index = self.tokenizer.pad_index
-        bsz = x.shape[0]
-
-        # Encode image(s) to latent vectors.
-        img_features = self.network.encode(x)
-
-        # Create a placeholder matrix for storing outputs from the network
-        indecies = torch.ones((bsz, self.max_output_len), dtype=torch.long).to(x.device)
-        indecies[:, 0] = start_index
-
-        for Sy in range(1, self.max_output_len):
-            tokens = indecies[:, :Sy]  # (B, Sy)
-            logits = self.network.decode(tokens, img_features)  # (B, C, Sy)
-            indecies_ = torch.argmax(logits, dim=1)  # (B, Sy)
-            indecies[:, Sy : Sy + 1] = indecies_[:, -1:]
-
-            # Early stopping of prediction loop if token is end or padding token.
-            if (
-                (indecies[:, Sy - 1] == end_index) | (indecies[:, Sy - 1] == pad_index)
-            ).all():
-                break
-
-        # Set all tokens after end token to pad token.
-        for Sy in range(1, self.max_output_len):
-            idx = (indecies[:, Sy - 1] == end_index) | (
-                indecies[:, Sy - 1] == pad_index
-            )
-            indecies[idx, Sy] = pad_index
-
-        return indecies
+        return self.decoder(x)
diff --git a/training/conf/config.yaml b/training/conf/config.yaml
index a95307f..e57a8a8 100644
--- a/training/conf/config.yaml
+++ b/training/conf/config.yaml
@@ -4,6 +4,7 @@ defaults:
   - _self_
   - callbacks: default
   - criterion: cross_entropy
+  - decoder: greedy
   - datamodule: iam_extended_paragraphs
   - hydra: default
   - logger: wandb
diff --git a/training/run.py b/training/run.py
index 429b1a2..288a1ef 100644
--- a/training/run.py
+++ b/training/run.py
@@ -1,5 +1,5 @@
 """Script to run experiments."""
-from typing import List, Optional, Type
+from typing import Callable, List, Optional, Type
 
 import hydra
 from loguru import logger as log
@@ -34,11 +34,19 @@ def run(config: DictConfig) -> Optional[float]:
     log.info(f"Instantiating criterion <{config.criterion._target_}>")
     loss_fn: Type[nn.Module] = hydra.utils.instantiate(config.criterion)
 
+    log.info(f"Instantiating decoder <{config.criterion._target_}>")
+    decoder: Type[Callable] = hydra.utils.instantiate(
+        config.decoder,
+        network=network,
+        tokenizer=datamodule.tokenizer,
+    )
+
     log.info(f"Instantiating model <{config.model._target_}>")
     model: LightningModule = hydra.utils.instantiate(
         config.model,
         network=network,
         tokenizer=datamodule.tokenizer,
+        decoder=decoder,
         loss_fn=loss_fn,
         optimizer_config=config.optimizer,
         lr_scheduler_config=config.lr_scheduler,