Trying to get the CNNTransformer to work, but it is hard.

12 files changed, 197 insertions, 47 deletions

diff --git a/src/text_recognizer/networks/__init__.py b/src/text_recognizer/networks/__init__.py
index 8b87797..6d88768 100644
--- a/src/text_recognizer/networks/__init__.py
+++ b/src/text_recognizer/networks/__init__.py
@@ -1,5 +1,6 @@
 """Network modules."""
 from .cnn_transformer import CNNTransformer
+from .cnn_transformer_encoder import CNNTransformerEncoder
 from .crnn import ConvolutionalRecurrentNetwork
 from .ctc import greedy_decoder
 from .densenet import DenseNet
@@ -15,6 +16,7 @@ from .wide_resnet import WideResidualNetwork
 
 __all__ = [
     "CNNTransformer",
+    "CNNTransformerEncoder",
     "ConvolutionalRecurrentNetwork",
     "DenseNet",
     "EmbeddingLoss",
diff --git a/src/text_recognizer/networks/cnn_transformer.py b/src/text_recognizer/networks/cnn_transformer.py
index 8666f11..3da2c9f 100644
--- a/src/text_recognizer/networks/cnn_transformer.py
+++ b/src/text_recognizer/networks/cnn_transformer.py
@@ -1,8 +1,7 @@
 """A DETR style transfomers but for text recognition."""
-from typing import Dict, Optional, Tuple, Type
+from typing import Dict, Optional, Tuple
 
-from einops.layers.torch import Rearrange
-from loguru import logger
+from einops import rearrange
 import torch
 from torch import nn
 from torch import Tensor
@@ -21,23 +20,32 @@ class CNNTransformer(nn.Module):
         hidden_dim: int,
         vocab_size: int,
         num_heads: int,
-        max_len: int,
+        adaptive_pool_dim: Tuple,
         expansion_dim: int,
         dropout_rate: float,
         trg_pad_index: int,
         backbone: str,
+        out_channels: int,
+        max_len: int,
         backbone_args: Optional[Dict] = None,
         activation: str = "gelu",
     ) -> None:
         super().__init__()
         self.trg_pad_index = trg_pad_index
-        self.backbone_args = backbone_args
+
         self.backbone = configure_backbone(backbone, backbone_args)
         self.character_embedding = nn.Embedding(vocab_size, hidden_dim)
-        self.position_encoding = PositionalEncoding(hidden_dim, dropout_rate, max_len)
-        self.collapse_spatial_dim = nn.Sequential(
-            Rearrange("b t h w -> b t (h w)"), nn.AdaptiveAvgPool2d((None, hidden_dim))
+
+        # self.conv = nn.Conv2d(out_channels, max_len, kernel_size=1)
+
+        self.position_encoding = PositionalEncoding(hidden_dim, dropout_rate)
+        self.row_embed = nn.Parameter(torch.rand(max_len, max_len // 2))
+        self.col_embed = nn.Parameter(torch.rand(max_len, max_len // 2))
+
+        self.adaptive_pool = (
+            nn.AdaptiveAvgPool2d((adaptive_pool_dim)) if adaptive_pool_dim else None
         )
+
         self.transformer = Transformer(
             num_encoder_layers,
             num_decoder_layers,
@@ -47,7 +55,8 @@ class CNNTransformer(nn.Module):
             dropout_rate,
             activation,
         )
-        self.head = nn.Linear(hidden_dim, vocab_size)
+
+        self.head = nn.Sequential(nn.Linear(hidden_dim, vocab_size),)
 
     def _create_trg_mask(self, trg: Tensor) -> Tensor:
         # Move this outside the transformer.
@@ -83,8 +92,22 @@ class CNNTransformer(nn.Module):
         if len(src.shape) < 4:
             src = src[(None,) * (4 - len(src.shape))]
         src = self.backbone(src)
-        src = self.collapse_spatial_dim(src)
-        src = self.position_encoding(src)
+        # src = self.conv(src)
+        if self.adaptive_pool is not None:
+            src = self.adaptive_pool(src)
+        H, W = src.shape[-2:]
+        src = rearrange(src, "b t h w -> b t (h w)")
+
+        # construct positional encodings
+        pos = torch.cat(
+            [
+                self.col_embed[:W].unsqueeze(0).repeat(H, 1, 1),
+                self.row_embed[:H].unsqueeze(1).repeat(1, W, 1),
+            ],
+            dim=-1,
+        ).unsqueeze(0)
+        pos = rearrange(pos, "b h w l -> b l (h w)")
+        src = pos + 0.1 * src
         return src
 
     def preprocess_target(self, trg: Tensor) -> Tuple[Tensor, Tensor]:
@@ -97,15 +120,16 @@ class CNNTransformer(nn.Module):
             Tuple[Tensor, Tensor]: Encoded target tensor and target mask.
 
         """
-        trg_mask = self._create_trg_mask(trg)
         trg = self.character_embedding(trg.long())
         trg = self.position_encoding(trg)
-        return trg, trg_mask
+        return trg
 
-    def forward(self, x: Tensor, trg: Tensor) -> Tensor:
+    def forward(self, x: Tensor, trg: Optional[Tensor] = None) -> Tensor:
         """Forward pass with CNN transfomer."""
-        src = self.preprocess_input(x)
-        trg, trg_mask = self.preprocess_target(trg)
-        out = self.transformer(src, trg, trg_mask=trg_mask)
+        h = self.preprocess_input(x)
+        trg_mask = self._create_trg_mask(trg)
+        trg = self.preprocess_target(trg)
+        out = self.transformer(h, trg, trg_mask=trg_mask)
+
         logits = self.head(out)
         return logits
diff --git a/src/text_recognizer/networks/cnn_transformer_encoder.py b/src/text_recognizer/networks/cnn_transformer_encoder.py
new file mode 100644
index 0000000..93626bf
--- /dev/null
+++ b/src/text_recognizer/networks/cnn_transformer_encoder.py
@@ -0,0 +1,73 @@
+"""Network with a CNN backend and a transformer encoder head."""
+from typing import Dict
+
+from einops import rearrange
+import torch
+from torch import nn
+from torch import Tensor
+
+from text_recognizer.networks.transformer import PositionalEncoding
+from text_recognizer.networks.util import configure_backbone
+
+
+class CNNTransformerEncoder(nn.Module):
+    """A CNN backbone with Transformer Encoder frontend for sequence prediction."""
+
+    def __init__(
+        self,
+        backbone: str,
+        backbone_args: Dict,
+        mlp_dim: int,
+        d_model: int,
+        nhead: int = 8,
+        dropout_rate: float = 0.1,
+        activation: str = "relu",
+        num_layers: int = 6,
+        num_classes: int = 80,
+        num_channels: int = 256,
+        max_len: int = 97,
+    ) -> None:
+        super().__init__()
+        self.d_model = d_model
+        self.nhead = nhead
+        self.dropout_rate = dropout_rate
+        self.activation = activation
+        self.num_layers = num_layers
+
+        self.backbone = configure_backbone(backbone, backbone_args)
+        self.position_encoding = PositionalEncoding(d_model, dropout_rate)
+        self.encoder = self._configure_encoder()
+
+        self.conv = nn.Conv2d(num_channels, max_len, kernel_size=1)
+
+        self.mlp = nn.Linear(mlp_dim, d_model)
+
+        self.head = nn.Linear(d_model, num_classes)
+
+    def _configure_encoder(self) -> nn.TransformerEncoder:
+        encoder_layer = nn.TransformerEncoderLayer(
+            d_model=self.d_model,
+            nhead=self.nhead,
+            dropout=self.dropout_rate,
+            activation=self.activation,
+        )
+        norm = nn.LayerNorm(self.d_model)
+        return nn.TransformerEncoder(
+            encoder_layer=encoder_layer, num_layers=self.num_layers, norm=norm
+        )
+
+    def forward(self, x: Tensor, targets: Tensor = None) -> Tensor:
+        """Forward pass through the network."""
+        if len(x.shape) < 4:
+            x = x[(None,) * (4 - len(x.shape))]
+
+        x = self.conv(self.backbone(x))
+        x = rearrange(x, "b c h w -> b c (h w)")
+        x = self.mlp(x)
+        x = self.position_encoding(x)
+        x = rearrange(x, "b c h-> c b h")
+        x = self.encoder(x)
+        x = rearrange(x, "c b h-> b c h")
+        logits = self.head(x)
+
+        return logits
diff --git a/src/text_recognizer/networks/crnn.py b/src/text_recognizer/networks/crnn.py
index 3e605e2..9747429 100644
--- a/src/text_recognizer/networks/crnn.py
+++ b/src/text_recognizer/networks/crnn.py
@@ -1,12 +1,9 @@
 """LSTM with CTC for handwritten text recognition within a line."""
-import importlib
-from pathlib import Path
-from typing import Callable, Dict, List, Optional, Tuple, Type, Union
+from typing import Dict, Tuple
 
 from einops import rearrange, reduce
-from einops.layers.torch import Rearrange, Reduce
+from einops.layers.torch import Rearrange
 from loguru import logger
-import torch
 from torch import nn
 from torch import Tensor
 
@@ -28,16 +25,21 @@ class ConvolutionalRecurrentNetwork(nn.Module):
         patch_size: Tuple[int, int] = (28, 28),
         stride: Tuple[int, int] = (1, 14),
         recurrent_cell: str = "lstm",
+        avg_pool: bool = False,
+        use_sliding_window: bool = True,
     ) -> None:
         super().__init__()
         self.backbone_args = backbone_args or {}
         self.patch_size = patch_size
         self.stride = stride
-        self.sliding_window = self._configure_sliding_window()
+        self.sliding_window = (
+            self._configure_sliding_window() if use_sliding_window else None
+        )
         self.input_size = input_size
         self.hidden_size = hidden_size
         self.backbone = configure_backbone(backbone, backbone_args)
         self.bidirectional = bidirectional
+        self.avg_pool = avg_pool
 
         if recurrent_cell.upper() in ["LSTM", "GRU"]:
             recurrent_cell = getattr(nn, recurrent_cell)
@@ -76,15 +78,27 @@ class ConvolutionalRecurrentNetwork(nn.Module):
         """Converts images to sequence of patches, feeds them to a CNN, then predictions are made with an LSTM."""
         if len(x.shape) < 4:
             x = x[(None,) * (4 - len(x.shape))]
-        x = self.sliding_window(x)
 
-        # Rearrange from a sequence of patches for feedforward network.
-        b, t = x.shape[:2]
-        x = rearrange(x, "b t c h w -> (b t) c h w", b=b, t=t)
-        x = self.backbone(x)
+        if self.sliding_window is not None:
+            # Create image patches with a sliding window kernel.
+            x = self.sliding_window(x)
+
+            # Rearrange from a sequence of patches for feedforward network.
+            b, t = x.shape[:2]
+            x = rearrange(x, "b t c h w -> (b t) c h w", b=b, t=t)
 
-        # Avgerage pooling.
-        x = reduce(x, "(b t) c h w -> t b c", "mean", b=b, t=t)
+            x = self.backbone(x)
+
+            # Avgerage pooling.
+            if self.avg_pool:
+                x = reduce(x, "(b t) c h w -> t b c", "mean", b=b, t=t)
+            else:
+                x = rearrange(x, "(b t) h -> t b h", b=b, t=t)
+        else:
+            # Encode the entire image with a CNN, and use the channels as temporal dimension.
+            b = x.shape[0]
+            x = self.backbone(x)
+            x = rearrange(x, "b c h w -> c b (h w)", b=b)
 
         # Sequence predictions.
         x, _ = self.rnn(x)
diff --git a/src/text_recognizer/networks/ctc.py b/src/text_recognizer/networks/ctc.py
index 2493d5c..af9b700 100644
--- a/src/text_recognizer/networks/ctc.py
+++ b/src/text_recognizer/networks/ctc.py
@@ -33,7 +33,7 @@ def greedy_decoder(
     """
 
     if character_mapper is None:
-        character_mapper = EmnistMapper()
+        character_mapper = EmnistMapper(pad_token="_")  # noqa: S106
 
     predictions = rearrange(torch.argmax(predictions, dim=2), "t b -> b t")
     decoded_predictions = []
diff --git a/src/text_recognizer/networks/densenet.py b/src/text_recognizer/networks/densenet.py
index d2aad60..7dc58d9 100644
--- a/src/text_recognizer/networks/densenet.py
+++ b/src/text_recognizer/networks/densenet.py
@@ -72,7 +72,7 @@ class _DenseBlock(nn.Module):
     ) -> None:
         super().__init__()
         self.dense_block = self._build_dense_blocks(
-            num_layers, in_channels, bn_size, growth_rate, dropout_rate, activation
+            num_layers, in_channels, bn_size, growth_rate, dropout_rate, activation,
         )
 
     def _build_dense_blocks(
@@ -219,7 +219,7 @@ class DenseNet(nn.Module):
 
     def forward(self, x: Tensor) -> Tensor:
         """Forward pass of Densenet."""
-        # If batch dimenstion is missing, it needs to be added.
+        # If batch dimenstion is missing, it will be added.
         if len(x.shape) < 4:
             x = x[(None,) * (4 - len(x.shape))]
         return self.densenet(x)
diff --git a/src/text_recognizer/networks/loss.py b/src/text_recognizer/networks/loss.py
index ff843cf..cf9fa0d 100644
--- a/src/text_recognizer/networks/loss.py
+++ b/src/text_recognizer/networks/loss.py
@@ -1,10 +1,12 @@
 """Implementations of custom loss functions."""
 from pytorch_metric_learning import distances, losses, miners, reducers
+import torch
 from torch import nn
 from torch import Tensor
+from torch.autograd import Variable
+import torch.nn.functional as F
 
-
-__all__ = ["EmbeddingLoss"]
+__all__ = ["EmbeddingLoss", "LabelSmoothingCrossEntropy"]
 
 
 class EmbeddingLoss:
@@ -32,3 +34,36 @@ class EmbeddingLoss:
         hard_pairs = self.miner(embeddings, labels)
         loss = self.loss_fn(embeddings, labels, hard_pairs)
         return loss
+
+
+class LabelSmoothingCrossEntropy(nn.Module):
+    """Label smoothing loss function."""
+
+    def __init__(
+        self,
+        classes: int,
+        smoothing: float = 0.0,
+        ignore_index: int = None,
+        dim: int = -1,
+    ) -> None:
+        super().__init__()
+        self.confidence = 1.0 - smoothing
+        self.smoothing = smoothing
+        self.ignore_index = ignore_index
+        self.cls = classes
+        self.dim = dim
+
+    def forward(self, pred: Tensor, target: Tensor) -> Tensor:
+        """Calculates the loss."""
+        pred = pred.log_softmax(dim=self.dim)
+        with torch.no_grad():
+            # true_dist = pred.data.clone()
+            true_dist = torch.zeros_like(pred)
+            true_dist.fill_(self.smoothing / (self.cls - 1))
+            true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
+            if self.ignore_index is not None:
+                true_dist[:, self.ignore_index] = 0
+                mask = torch.nonzero(target == self.ignore_index, as_tuple=False)
+                if mask.dim() > 0:
+                    true_dist.index_fill_(0, mask.squeeze(), 0.0)
+        return torch.mean(torch.sum(-true_dist * pred, dim=self.dim))
diff --git a/src/text_recognizer/networks/transformer/positional_encoding.py b/src/text_recognizer/networks/transformer/positional_encoding.py
index a47141b..1ba5537 100644
--- a/src/text_recognizer/networks/transformer/positional_encoding.py
+++ b/src/text_recognizer/networks/transformer/positional_encoding.py
@@ -13,6 +13,7 @@ class PositionalEncoding(nn.Module):
     ) -> None:
         super().__init__()
         self.dropout = nn.Dropout(p=dropout_rate)
+        self.max_len = max_len
 
         pe = torch.zeros(max_len, hidden_dim)
         position = torch.arange(0, max_len).unsqueeze(1)
diff --git a/src/text_recognizer/networks/transformer/sparse_transformer.py b/src/text_recognizer/networks/transformer/sparse_transformer.py
deleted file mode 100644
index 8c391c8..0000000
--- a/src/text_recognizer/networks/transformer/sparse_transformer.py
+++ /dev/null
@@ -1 +0,0 @@
-"""Encoder and Decoder modules using spares activations."""
diff --git a/src/text_recognizer/networks/transformer/transformer.py b/src/text_recognizer/networks/transformer/transformer.py
index 1c9c7dd..c6e943e 100644
--- a/src/text_recognizer/networks/transformer/transformer.py
+++ b/src/text_recognizer/networks/transformer/transformer.py
@@ -230,6 +230,7 @@ class Transformer(nn.Module):
     ) -> Tensor:
         """Forward pass through the transformer."""
         if src.shape[0] != trg.shape[0]:
+            print(trg.shape)
             raise RuntimeError("The batch size of the src and trg must be the same.")
         if src.shape[2] != trg.shape[2]:
             raise RuntimeError(
diff --git a/src/text_recognizer/networks/util.py b/src/text_recognizer/networks/util.py
index 0d08506..b31e640 100644
--- a/src/text_recognizer/networks/util.py
+++ b/src/text_recognizer/networks/util.py
@@ -28,7 +28,7 @@ def sliding_window(
     c = images.shape[1]
     patches = unfold(images)
     patches = rearrange(
-        patches, "b (c h w) t -> b t c h w", c=c, h=patch_size[0], w=patch_size[1]
+        patches, "b (c h w) t -> b t c h w", c=c, h=patch_size[0], w=patch_size[1],
     )
     return patches
 
@@ -77,7 +77,7 @@ def configure_backbone(backbone: str, backbone_args: Dict) -> Type[nn.Module]:
 
     if "remove_layers" in backbone_args and backbone_args["remove_layers"] is not None:
         backbone = nn.Sequential(
-            *list(backbone.children())[0][: -backbone_args["remove_layers"]]
+            *list(backbone.children())[:][: -backbone_args["remove_layers"]]
         )
 
     return backbone
diff --git a/src/text_recognizer/networks/vision_transformer.py b/src/text_recognizer/networks/vision_transformer.py
index 4d204d3..f227954 100644
--- a/src/text_recognizer/networks/vision_transformer.py
+++ b/src/text_recognizer/networks/vision_transformer.py
@@ -29,9 +29,9 @@ class VisionTransformer(nn.Module):
         num_heads: int,
         max_len: int,
         expansion_dim: int,
-        mlp_dim: int,
         dropout_rate: float,
         trg_pad_index: int,
+        mlp_dim: Optional[int] = None,
         patch_size: Tuple[int, int] = (28, 28),
         stride: Tuple[int, int] = (1, 14),
         activation: str = "gelu",
@@ -46,6 +46,7 @@ class VisionTransformer(nn.Module):
         self.slidning_window = self._configure_sliding_window()
         self.character_embedding = nn.Embedding(vocab_size, hidden_dim)
         self.position_encoding = PositionalEncoding(hidden_dim, dropout_rate, max_len)
+        self.mlp_dim = mlp_dim
 
         self.use_backbone = False
         if backbone is None:
@@ -54,6 +55,8 @@ class VisionTransformer(nn.Module):
             )
         else:
             self.backbone = configure_backbone(backbone, backbone_args)
+            if mlp_dim:
+                self.mlp = nn.Linear(mlp_dim, hidden_dim)
             self.use_backbone = True
 
         self.transformer = Transformer(
@@ -66,13 +69,7 @@ class VisionTransformer(nn.Module):
             activation,
         )
 
-        self.head = nn.Sequential(
-            nn.LayerNorm(hidden_dim),
-            nn.Linear(hidden_dim, mlp_dim),
-            nn.GELU(),
-            nn.Dropout(p=dropout_rate),
-            nn.Linear(mlp_dim, vocab_size),
-        )
+        self.head = nn.Sequential(nn.Linear(hidden_dim, vocab_size),)
 
     def _configure_sliding_window(self) -> nn.Sequential:
         return nn.Sequential(
@@ -110,7 +107,11 @@ class VisionTransformer(nn.Module):
         if self.use_backbone:
             x = rearrange(x, "b t c h w -> (b t) c h w", b=b, t=t)
             x = self.backbone(x)
-            x = rearrange(x, "(b t) h -> b t h", b=b, t=t)
+            if self.mlp_dim:
+                x = rearrange(x, "(b t) c h w -> b t (c h w)", b=b, t=t)
+                x = self.mlp(x)
+            else:
+                x = rearrange(x, "(b t) h -> b t h", b=b, t=t)
         else:
             x = rearrange(x, "b t c h w -> b t (c h w)", b=b, t=t)
             x = self.linear_projection(x)