14 files changed, 227 insertions, 170 deletions

diff --git a/text_recognizer/networks/__init__.py b/text_recognizer/networks/__init__.py
index 41fd43f..63b43b2 100644
--- a/text_recognizer/networks/__init__.py
+++ b/text_recognizer/networks/__init__.py
@@ -1,2 +1,4 @@
 """Network modules"""
+from .backbones import EfficientNet
 from .vqvae import VQVAE
+from .cnn_transformer import CNNTransformer
diff --git a/text_recognizer/networks/backbones/__init__.py b/text_recognizer/networks/backbones/__init__.py
new file mode 100644
index 0000000..25aed0e
--- /dev/null
+++ b/text_recognizer/networks/backbones/__init__.py
@@ -0,0 +1,2 @@
+"""Vision backbones."""
+from .efficientnet import EfficientNet
diff --git a/text_recognizer/networks/backbones/efficientnet.py b/text_recognizer/networks/backbones/efficientnet.py
new file mode 100644
index 0000000..61dea77
--- /dev/null
+++ b/text_recognizer/networks/backbones/efficientnet.py
@@ -0,0 +1,145 @@
+"""Efficient net b0 implementation."""
+import torch
+from torch import nn
+from torch import Tensor
+
+
+class ConvNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int,
+        padding: int,
+        groups: int = 1,
+    ) -> None:
+        super().__init__()
+        self.block = nn.Sequential(
+            nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=kernel_size,
+                stride=stride,
+                padding=padding,
+                groups=groups,
+                bias=False,
+            ),
+            nn.BatchNorm2d(num_features=out_channels),
+            nn.SiLU(inplace=True),
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.block(x)
+
+
+class SqueezeExcite(nn.Module):
+    def __init__(self, in_channels: int, reduce_dim: int) -> None:
+        super().__init__()
+        self.se = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),  # [C, H, W] -> [C, 1, 1]
+            nn.Conv2d(in_channels=in_channels, out_channels=reduce_dim, kernel_size=1),
+            nn.SiLU(),
+            nn.Conv2d(in_channels=reduce_dim, out_channels=in_channels, kernel_size=1),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x * self.se(x)
+
+
+class InvertedResidulaBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int,
+        padding: int,
+        expand_ratio: float,
+        reduction: int = 4,
+        survival_prob: float = 0.8,
+    ) -> None:
+        super().__init__()
+        self.survival_prob = survival_prob
+        self.use_residual = in_channels == out_channels and stride == 1
+        hidden_dim = in_channels * expand_ratio
+        self.expand = in_channels != hidden_dim
+        reduce_dim = in_channels // reduction
+
+        if self.expand:
+            self.expand_conv = ConvNorm(
+                in_channels, hidden_dim, kernel_size=3, stride=1, padding=1
+            )
+
+        self.conv = nn.Sequential(
+            ConvNorm(
+                hidden_dim, hidden_dim, kernel_size, stride, padding, groups=hidden_dim
+            ),
+            SqueezeExcite(hidden_dim, reduce_dim),
+            nn.Conv2d(
+                in_channels=hidden_dim,
+                out_channels=out_channels,
+                kernel_size=1,
+                bias=False,
+            ),
+            nn.BatchNorm2d(num_features=out_channels),
+        )
+
+    def stochastic_depth(self, x: Tensor) -> Tensor:
+        if not self.training:
+            return x
+
+        binary_tensor = (
+            torch.rand(x.shape[0], 1, 1, 1, device=x.device) < self.survival_prob
+        )
+        return torch.div(x, self.survival_prob) * binary_tensor
+
+    def forward(self, x: Tensor) -> Tensor:
+        out = self.expand_conv(x) if self.expand else x
+        if self.use_residual:
+            return self.stochastic_depth(self.conv(out)) + x
+        return self.conv(out)
+
+
+class EfficientNet(nn.Module):
+    """Efficient net b0 backbone."""
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.base_model = [
+            # expand_ratio, channels, repeats, stride, kernel_size
+            [1, 16, 1, 1, 3],
+            [6, 24, 2, 2, 3],
+            [6, 40, 2, 2, 5],
+            [6, 80, 3, 2, 3],
+            [6, 112, 3, 1, 5],
+            [6, 192, 4, 2, 5],
+            [6, 320, 1, 1, 3],
+        ]
+
+        self.backbone = self._build_b0()
+
+    def _build_b0(self) -> nn.Sequential:
+        in_channels = 32
+        layers = [ConvNorm(1, in_channels, 3, stride=2, padding=1)]
+
+        for expand_ratio, out_channels, repeats, stride, kernel_size in self.base_model:
+            for i in range(repeats):
+                layers.append(
+                    InvertedResidulaBlock(
+                        in_channels,
+                        out_channels,
+                        expand_ratio=expand_ratio,
+                        stride=stride if i == 0 else 1,
+                        kernel_size=kernel_size,
+                        padding=kernel_size // 2,
+                    )
+                )
+                in_channels = out_channels
+        layers.append(ConvNorm(in_channels, 256, kernel_size=1, stride=1, padding=0))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.backbone(x)
diff --git a/text_recognizer/networks/cnn_transformer.py b/text_recognizer/networks/cnn_transformer.py
index e23a15d..d42c29d 100644
--- a/text_recognizer/networks/cnn_transformer.py
+++ b/text_recognizer/networks/cnn_transformer.py
@@ -33,8 +33,8 @@ NUM_WORD_PIECES = 1000
 class CNNTransformer(nn.Module):
     def __init__(
         self,
-        input_shape: Sequence[int],
-        output_shape: Sequence[int],
+        input_dim: Sequence[int],
+        output_dims: Sequence[int],
         encoder: Union[DictConfig, Dict],
         vocab_size: Optional[int] = None,
         num_decoder_layers: int = 4,
@@ -43,22 +43,29 @@ class CNNTransformer(nn.Module):
         expansion_dim: int = 1024,
         dropout_rate: float = 0.1,
         transformer_activation: str = "glu",
+        *args,
+        **kwargs,
     ) -> None:
+        super().__init__()
         self.vocab_size = (
             NUM_WORD_PIECES + NUM_SPECIAL_TOKENS if vocab_size is None else vocab_size
         )
+        self.pad_index = 3  # TODO: fix me
         self.hidden_dim = hidden_dim
-        self.max_output_length = output_shape[0]
+        self.max_output_length = output_dims[0]
 
         # Image backbone
         self.encoder = self._configure_encoder(encoder)
+        self.encoder_proj = nn.Conv2d(256, hidden_dim, kernel_size=1)
         self.feature_map_encoding = PositionalEncoding2D(
-            hidden_dim=hidden_dim, max_h=input_shape[1], max_w=input_shape[2]
+            hidden_dim=hidden_dim, max_h=input_dim[1], max_w=input_dim[2]
         )
 
         # Target token embedding
         self.trg_embedding = nn.Embedding(self.vocab_size, hidden_dim)
-        self.trg_position_encoding = PositionalEncoding(hidden_dim, dropout_rate)
+        self.trg_position_encoding = PositionalEncoding(
+            hidden_dim, dropout_rate, max_len=output_dims[0]
+        )
 
         # Transformer decoder
         self.decoder = Decoder(
@@ -86,24 +93,25 @@ class CNNTransformer(nn.Module):
         self.head.weight.data.uniform_(-0.1, 0.1)
 
         nn.init.kaiming_normal_(
-            self.feature_map_encoding.weight.data,
+            self.encoder_proj.weight.data,
             a=0,
             mode="fan_out",
             nonlinearity="relu",
         )
-        if self.feature_map_encoding.bias is not None:
+        if self.encoder_proj.bias is not None:
             _, fan_out = nn.init._calculate_fan_in_and_fan_out(
-                self.feature_map_encoding.weight.data
+                self.encoder_proj.weight.data
             )
             bound = 1 / math.sqrt(fan_out)
-            nn.init.normal_(self.feature_map_encoding.bias, -bound, bound)
+            nn.init.normal_(self.encoder_proj.bias, -bound, bound)
 
     @staticmethod
     def _configure_encoder(encoder: Union[DictConfig, Dict]) -> Type[nn.Module]:
         encoder = OmegaConf.create(encoder)
+        args = encoder.args or {}
         network_module = importlib.import_module("text_recognizer.networks")
         encoder_class = getattr(network_module, encoder.type)
-        return encoder_class(**encoder.args)
+        return encoder_class(**args)
 
     def encode(self, image: Tensor) -> Tensor:
         """Extracts image features with backbone.
@@ -121,6 +129,7 @@ class CNNTransformer(nn.Module):
         """
         # Extract image features.
         image_features = self.encoder(image)
+        image_features = self.encoder_proj(image_features)
 
         # Add 2d encoding to the feature maps.
         image_features = self.feature_map_encoding(image_features)
@@ -133,11 +142,19 @@ class CNNTransformer(nn.Module):
         """Decodes image features with transformer decoder."""
         trg_mask = target_padding_mask(trg=trg, pad_index=self.pad_index)
         trg = self.trg_embedding(trg) * math.sqrt(self.hidden_dim)
+        trg = rearrange(trg, "b t d -> t b d")
         trg = self.trg_position_encoding(trg)
+        trg = rearrange(trg, "t b d -> b t d")
         out = self.decoder(trg=trg, memory=memory, trg_mask=trg_mask, memory_mask=None)
         logits = self.head(out)
         return logits
 
+    def forward(self, image: Tensor, trg: Tensor) -> Tensor:
+        image_features = self.encode(image)
+        output = self.decode(image_features, trg)
+        output = rearrange(output, "b t c -> b c t")
+        return output
+
     def predict(self, image: Tensor) -> Tensor:
         """Transcribes text in image(s)."""
         bsz = image.shape[0]
diff --git a/text_recognizer/networks/coat/__init__.py b/text_recognizer/networks/coat/__init__.py
deleted file mode 100644
index e69de29..0000000
--- a/text_recognizer/networks/coat/__init__.py
+++ /dev/null
diff --git a/text_recognizer/networks/coat/factor_attention.py b/text_recognizer/networks/coat/factor_attention.py
deleted file mode 100644
index f91c5ef..0000000
--- a/text_recognizer/networks/coat/factor_attention.py
+++ /dev/null
@@ -1,9 +0,0 @@
-"""Factorized attention with convolutional relative positional encodings."""
-from torch import nn
-
-
-class FactorAttention(nn.Module):
-    """Factorized attention with relative positional encodings."""
-    def __init__(self, dim: int, num_heads: int) -> None:
-        pass
-
diff --git a/text_recognizer/networks/coat/patch_embedding.py b/text_recognizer/networks/coat/patch_embedding.py
deleted file mode 100644
index 3b7b76a..0000000
--- a/text_recognizer/networks/coat/patch_embedding.py
+++ /dev/null
@@ -1,38 +0,0 @@
-"""Patch embedding for images and feature maps."""
-from typing import Sequence, Tuple
-
-from einops import rearrange
-from loguru import logger
-from torch import nn
-from torch import Tensor
-
-
-class PatchEmbedding(nn.Module):
-    """Patch embedding of images."""
-
-    def __init__(
-        self,
-        image_shape: Sequence[int],
-        patch_size: int = 16,
-        in_channels: int = 1,
-        embedding_dim: int = 512,
-    ) -> None:
-        if image_shape[0] % patch_size == 0 and image_shape[1] % patch_size == 0:
-            logger.error(
-                f"Image shape {image_shape} not divisable by patch size {patch_size}"
-            )
-
-        self.patch_size = patch_size
-        self.embedding = nn.Conv2d(
-            in_channels, embedding_dim, kernel_size=patch_size, stride=patch_size
-        )
-        self.norm = nn.LayerNorm(embedding_dim)
-
-    def forward(self, x: Tensor) -> Tuple[Tensor, Tuple[int, int]]:
-        """Embeds image or feature maps with patch embedding."""
-        _, _, h, w = x.shape
-        h_out, w_out = h // self.patch_size, w // self.patch_size
-        x = self.embedding(x)
-        x = rearrange(x, "b c h w -> b (h w) c")
-        x = self.norm(x)
-        return x, (h_out, w_out)
diff --git a/text_recognizer/networks/coat/positional_encodings.py b/text_recognizer/networks/coat/positional_encodings.py
deleted file mode 100644
index 925db04..0000000
--- a/text_recognizer/networks/coat/positional_encodings.py
+++ /dev/null
@@ -1,76 +0,0 @@
-"""Positional encodings for input sequence to transformer."""
-from typing import Dict, Union, Tuple
-
-from einops import rearrange
-from loguru import logger
-import torch
-from torch import nn
-from torch import Tensor
-
-
-class RelativeEncoding(nn.Module):
-    """Relative positional encoding."""
-    def __init__(self, channels: int, heads: int, windows: Union[int, Dict[int, int]]) -> None:
-        super().__init__()
-        self.windows = {windows: heads} if isinstance(windows, int) else windows
-        self.heads = list(self.windows.values())
-        self.channel_heads = [head * channels for head in self.heads]
-        self.convs = nn.ModuleList([
-            nn.Conv2d(in_channels=head * channels,
-                             out_channels=head * channels,
-                             kernel_shape=window,
-                             padding=window // 2,
-                             dilation=1,
-                             groups=head * channels,
-            ) for window, head in self.windows.items()])
-
-    def forward(self, q: Tensor, v: Tensor, shape: Tuple[int, int]) -> Tensor:
-        """Applies relative positional encoding."""
-        b, heads, hw, c = q.shape
-        h, w = shape
-        if hw != h * w:
-            logger.exception(f"Query width {hw} neq to height x width {h * w}")
-            raise ValueError
-        
-        v = rearrange(v, "b heads (h w) c -> b (heads c) h w", h=h, w=w)
-        v = torch.split(v, self.channel_heads, dim=1)
-        v = [conv(x) for conv, x in zip(self.convs, v)]
-        v = torch.cat(v, dim=1)
-        v = rearrange(v, "b (heads c) h w -> b heads (h w) c", heads=heads)
-
-        encoding = q * v
-        zeros = torch.zeros((b, heads, 1, c), dtype=q.dtype, layout=q.layout, device=q.device)
-        encoding = torch.cat((zeros, encoding), dim=2)
-        return encoding
-
-
-class PositionalEncoding(nn.Module):
-    """Convolutional positional encoding."""
-    def __init__(self, dim: int, k: int = 3) -> None:
-        super().__init__()
-        self.encode = nn.Conv2d(in_channels=dim, out_channels=dim, kernel_size=k, stride=1, padding=k//2, groups=dim)
-
-    def forward(self, x: Tensor, shape: Tuple[int, int]) -> Tensor:
-        """Applies convolutional encoding."""
-        _, hw, _ = x.shape
-        h, w = shape
-
-        if hw != h * w:
-            logger.exception(f"Query width {hw} neq to height x width {h * w}")
-            raise ValueError
-
-        # Depthwise convolution.
-        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
-        x = self.encode(x) + x
-        x = rearrange(x, "b c h w -> b (h w) c")
-        return x
-
-
-        
-
-
-
-            
-            
-
-        
diff --git a/text_recognizer/networks/residual_network.py b/text_recognizer/networks/residual_network.py
index da7553d..c33f419 100644
--- a/text_recognizer/networks/residual_network.py
+++ b/text_recognizer/networks/residual_network.py
@@ -20,11 +20,7 @@ class Conv2dAuto(nn.Conv2d):
 
 def conv_bn(in_channels: int, out_channels: int, *args, **kwargs) -> nn.Sequential:
     """3x3 convolution with batch norm."""
-    conv3x3 = partial(
-        Conv2dAuto,
-        kernel_size=3,
-        bias=False,
-    )
+    conv3x3 = partial(Conv2dAuto, kernel_size=3, bias=False,)
     return nn.Sequential(
         conv3x3(in_channels, out_channels, *args, **kwargs),
         nn.BatchNorm2d(out_channels),
diff --git a/text_recognizer/networks/transducer/transducer.py b/text_recognizer/networks/transducer/transducer.py
index b10f93a..d7e3d08 100644
--- a/text_recognizer/networks/transducer/transducer.py
+++ b/text_recognizer/networks/transducer/transducer.py
@@ -392,12 +392,7 @@ def load_transducer_loss(
         transitions = gtn.load(str(processed_path / transitions))
 
     preprocessor = Preprocessor(
-        data_dir,
-        num_features,
-        tokens_path,
-        lexicon_path,
-        use_words,
-        prepend_wordsep,
+        data_dir, num_features, tokens_path, lexicon_path, use_words, prepend_wordsep,
     )
 
     num_tokens = preprocessor.num_tokens
diff --git a/text_recognizer/networks/transformer/positional_encoding.py b/text_recognizer/networks/transformer/positional_encoding.py
index 5874e97..c50afc3 100644
--- a/text_recognizer/networks/transformer/positional_encoding.py
+++ b/text_recognizer/networks/transformer/positional_encoding.py
@@ -33,7 +33,10 @@ class PositionalEncoding(nn.Module):
 
     def forward(self, x: Tensor) -> Tensor:
         """Encodes the tensor with a postional embedding."""
-        x = x + self.pe[:, : x.shape[1]]
+        # [T, B, D]
+        if x.shape[2] != self.pe.shape[2]:
+            raise ValueError(f"x shape does not match pe in the 3rd dim.")
+        x = x + self.pe[: x.shape[0]]
         return self.dropout(x)
 
 
@@ -48,6 +51,7 @@ class PositionalEncoding2D(nn.Module):
         pe = self.make_pe(hidden_dim, max_h, max_w)
         self.register_buffer("pe", pe)
 
+    @staticmethod
     def make_pe(hidden_dim: int, max_h: int, max_w: int) -> Tensor:
         """Returns 2d postional encoding."""
         pe_h = PositionalEncoding.make_pe(
diff --git a/text_recognizer/networks/transformer/rotary_embedding.py b/text_recognizer/networks/transformer/rotary_embedding.py
new file mode 100644
index 0000000..5e80572
--- /dev/null
+++ b/text_recognizer/networks/transformer/rotary_embedding.py
@@ -0,0 +1,39 @@
+"""Roatary embedding.
+
+Stolen from lucidrains:
+    https://github.com/lucidrains/x-transformers/blob/main/x_transformers/x_transformers.py
+
+Explanation of roatary:
+    https://blog.eleuther.ai/rotary-embeddings/
+
+"""
+from typing import Tuple
+
+from einops import rearrange
+import torch
+from torch import nn
+from torch import Tensor
+
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+    def forward(self, x: Tensor, seq_dim: int = 1) -> Tensor:
+        t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq)
+        freqs = torch.einsum("i , j -> i j", t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        return emb[None, :, :]
+
+
+def rotate_half(x: Tensor) -> Tensor:
+    x = rearrange(x, "... (j d) -> ... j d", j=2)
+    x1, x2 = x.unbind(dim=-2)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q: Tensor, k: Tensor, freqs: Tensor) -> Tuple[Tensor, Tensor]:
+    q, k = map(lambda t: (t * freqs.cos()) + (rotate_half(t) * freqs.sin()), (q, k))
+    return q, k
diff --git a/text_recognizer/networks/vqvae/decoder.py b/text_recognizer/networks/vqvae/decoder.py
index 93a1e43..32de912 100644
--- a/text_recognizer/networks/vqvae/decoder.py
+++ b/text_recognizer/networks/vqvae/decoder.py
@@ -44,12 +44,7 @@ class Decoder(nn.Module):
 
         # Configure encoder.
         self.decoder = self._build_decoder(
-            channels,
-            kernel_sizes,
-            strides,
-            num_residual_layers,
-            activation,
-            dropout,
+            channels, kernel_sizes, strides, num_residual_layers, activation, dropout,
         )
 
     def _build_decompression_block(
@@ -78,9 +73,7 @@ class Decoder(nn.Module):
             )
 
             if self.upsampling and i < len(self.upsampling):
-                modules.append(
-                    nn.Upsample(size=self.upsampling[i]),
-                )
+                modules.append(nn.Upsample(size=self.upsampling[i]),)
 
             if dropout is not None:
                 modules.append(dropout)
@@ -109,12 +102,7 @@ class Decoder(nn.Module):
     ) -> nn.Sequential:
 
         self.res_block.append(
-            nn.Conv2d(
-                self.embedding_dim,
-                channels[0],
-                kernel_size=1,
-                stride=1,
-            )
+            nn.Conv2d(self.embedding_dim, channels[0], kernel_size=1, stride=1,)
         )
 
         # Bottleneck module.
diff --git a/text_recognizer/networks/vqvae/encoder.py b/text_recognizer/networks/vqvae/encoder.py
index b0cceed..65801df 100644
--- a/text_recognizer/networks/vqvae/encoder.py
+++ b/text_recognizer/networks/vqvae/encoder.py
@@ -11,10 +11,7 @@ from text_recognizer.networks.vqvae.vector_quantizer import VectorQuantizer
 
 class _ResidualBlock(nn.Module):
     def __init__(
-        self,
-        in_channels: int,
-        out_channels: int,
-        dropout: Optional[Type[nn.Module]],
+        self, in_channels: int, out_channels: int, dropout: Optional[Type[nn.Module]],
     ) -> None:
         super().__init__()
         self.block = [
@@ -138,12 +135,7 @@ class Encoder(nn.Module):
         )
 
         encoder.append(
-            nn.Conv2d(
-                channels[-1],
-                self.embedding_dim,
-                kernel_size=1,
-                stride=1,
-            )
+            nn.Conv2d(channels[-1], self.embedding_dim, kernel_size=1, stride=1,)
         )
 
         return nn.Sequential(*encoder)