Rename and add flash atten

20 files changed, 662 insertions, 0 deletions

diff --git a/text_recognizer/network/__init__.py b/text_recognizer/network/__init__.py
new file mode 100644
index 0000000..d9ef58b
--- /dev/null
+++ b/text_recognizer/network/__init__.py
@@ -0,0 +1 @@
+"""Network modules"""
diff --git a/text_recognizer/network/convnext/__init__.py b/text_recognizer/network/convnext/__init__.py
new file mode 100644
index 0000000..dcff3fc
--- /dev/null
+++ b/text_recognizer/network/convnext/__init__.py
@@ -0,0 +1,7 @@
+"""Convnext module."""
+from text_recognizer.network.convnext.attention import (
+    Attention,
+    FeedForward,
+    TransformerBlock,
+)
+from text_recognizer.network.convnext.convnext import ConvNext
diff --git a/text_recognizer/network/convnext/attention.py b/text_recognizer/network/convnext/attention.py
new file mode 100644
index 0000000..6bc9692
--- /dev/null
+++ b/text_recognizer/network/convnext/attention.py
@@ -0,0 +1,79 @@
+"""Convolution self attention block."""
+
+import torch.nn.functional as F
+from einops import rearrange
+from torch import Tensor, einsum, nn
+
+from text_recognizer.network.convnext.norm import LayerNorm
+from text_recognizer.network.convnext.residual import Residual
+
+
+def l2norm(t: Tensor) -> Tensor:
+    return F.normalize(t, dim=-1)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim: int, mult: int = 4) -> None:
+        super().__init__()
+        inner_dim = int(dim * mult)
+        self.fn = Residual(
+            nn.Sequential(
+                LayerNorm(dim),
+                nn.Conv2d(dim, inner_dim, 1, bias=False),
+                nn.GELU(),
+                LayerNorm(inner_dim),
+                nn.Conv2d(inner_dim, dim, 1, bias=False),
+            )
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.fn(x)
+
+
+class Attention(nn.Module):
+    def __init__(
+        self, dim: int, heads: int = 4, dim_head: int = 64, scale: int = 8
+    ) -> None:
+        super().__init__()
+        self.scale = scale
+        self.heads = heads
+        inner_dim = heads * dim_head
+        self.norm = LayerNorm(dim)
+
+        self.to_qkv = nn.Conv2d(dim, inner_dim * 3, 1, bias=False)
+        self.to_out = nn.Conv2d(inner_dim, dim, 1, bias=False)
+
+    def forward(self, x: Tensor) -> Tensor:
+        h, w = x.shape[-2:]
+
+        residual = x.clone()
+
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x).chunk(3, dim=1)
+        q, k, v = map(
+            lambda t: rearrange(t, "b (h c) ... -> b h (...) c", h=self.heads),
+            (q, k, v),
+        )
+
+        q, k = map(l2norm, (q, k))
+
+        sim = einsum("b h i d, b h j d -> b h i j", q, k) * self.scale
+        attn = sim.softmax(dim=-1)
+
+        out = einsum("b h i j, b h j d -> b h i d", attn, v)
+
+        out = rearrange(out, "b h (x y) d -> b (h d) x y", x=h, y=w)
+        return self.to_out(out) + residual
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, attn: Attention, ff: FeedForward) -> None:
+        super().__init__()
+        self.attn = attn
+        self.ff = ff
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.attn(x)
+        x = self.ff(x)
+        return x
diff --git a/text_recognizer/network/convnext/convnext.py b/text_recognizer/network/convnext/convnext.py
new file mode 100644
index 0000000..6acf059
--- /dev/null
+++ b/text_recognizer/network/convnext/convnext.py
@@ -0,0 +1,77 @@
+"""ConvNext module."""
+from typing import Optional, Sequence
+
+from torch import Tensor, nn
+
+from text_recognizer.network.convnext.attention import TransformerBlock
+from text_recognizer.network.convnext.downsample import Downsample
+from text_recognizer.network.convnext.norm import LayerNorm
+
+
+class ConvNextBlock(nn.Module):
+    """ConvNext block."""
+
+    def __init__(self, dim: int, dim_out: int, mult: int) -> None:
+        super().__init__()
+        self.ds_conv = nn.Conv2d(
+            dim, dim, kernel_size=(7, 7), padding="same", groups=dim
+        )
+        self.net = nn.Sequential(
+            LayerNorm(dim),
+            nn.Conv2d(dim, dim_out * mult, kernel_size=(3, 3), padding="same"),
+            nn.GELU(),
+            nn.Conv2d(dim_out * mult, dim_out, kernel_size=(3, 3), padding="same"),
+        )
+        self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity()
+
+    def forward(self, x: Tensor) -> Tensor:
+        h = self.ds_conv(x)
+        h = self.net(h)
+        return h + self.res_conv(x)
+
+
+class ConvNext(nn.Module):
+    def __init__(
+        self,
+        dim: int = 16,
+        dim_mults: Sequence[int] = (2, 4, 8),
+        depths: Sequence[int] = (3, 3, 6),
+        downsampling_factors: Sequence[Sequence[int]] = ((2, 2), (2, 2), (2, 2)),
+        attn: Optional[TransformerBlock] = None,
+    ) -> None:
+        super().__init__()
+        dims = (dim, *map(lambda m: m * dim, dim_mults))
+        self.attn = attn if attn is not None else nn.Identity()
+        self.out_channels = dims[-1]
+        self.stem = nn.Conv2d(1, dims[0], kernel_size=7, padding="same")
+        self.layers = nn.ModuleList([])
+
+        for i in range(len(dims) - 1):
+            dim_in, dim_out = dims[i], dims[i + 1]
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        ConvNextBlock(dim_in, dim_in, 2),
+                        nn.ModuleList(
+                            [ConvNextBlock(dim_in, dim_in, 2) for _ in range(depths[i])]
+                        ),
+                        Downsample(dim_in, dim_out, downsampling_factors[i]),
+                    ]
+                )
+            )
+        self.norm = LayerNorm(dims[-1])
+
+    def _init_weights(self, m):
+        if isinstance(m, (nn.Conv2d, nn.Linear)):
+            nn.init.trunc_normal_(m.weight, std=0.02)
+            nn.init.constant_(m.bias, 0)
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.stem(x)
+        for init_block, blocks, down in self.layers:
+            x = init_block(x)
+            for fn in blocks:
+                x = fn(x)
+            x = down(x)
+        x = self.attn(x)
+        return self.norm(x)
diff --git a/text_recognizer/network/convnext/downsample.py b/text_recognizer/network/convnext/downsample.py
new file mode 100644
index 0000000..a8a0466
--- /dev/null
+++ b/text_recognizer/network/convnext/downsample.py
@@ -0,0 +1,21 @@
+"""Convnext downsample module."""
+from typing import Tuple
+
+from einops.layers.torch import Rearrange
+from torch import Tensor, nn
+
+
+class Downsample(nn.Module):
+    """Downsamples feature maps by patches."""
+
+    def __init__(self, dim: int, dim_out: int, factors: Tuple[int, int]) -> None:
+        super().__init__()
+        s1, s2 = factors
+        self.fn = nn.Sequential(
+            Rearrange("b c (h s1) (w s2) -> b (c s1 s2) h w", s1=s1, s2=s2),
+            nn.Conv2d(dim * s1 * s2, dim_out, 1),
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies patch function."""
+        return self.fn(x)
diff --git a/text_recognizer/network/convnext/norm.py b/text_recognizer/network/convnext/norm.py
new file mode 100644
index 0000000..3355de9
--- /dev/null
+++ b/text_recognizer/network/convnext/norm.py
@@ -0,0 +1,18 @@
+"""Layer norm for conv layers."""
+import torch
+from torch import Tensor, nn
+
+
+class LayerNorm(nn.Module):
+    """Layer norm for convolutions."""
+
+    def __init__(self, dim: int) -> None:
+        super().__init__()
+        self.gamma = nn.Parameter(torch.ones(1, dim, 1, 1))
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies layer norm."""
+        eps = 1e-5 if x.dtype == torch.float32 else 1e-3
+        var = torch.var(x, dim=1, unbiased=False, keepdim=True)
+        mean = torch.mean(x, dim=1, keepdim=True)
+        return (x - mean) / (var + eps).sqrt() * self.gamma
diff --git a/text_recognizer/network/convnext/residual.py b/text_recognizer/network/convnext/residual.py
new file mode 100644
index 0000000..dfc2847
--- /dev/null
+++ b/text_recognizer/network/convnext/residual.py
@@ -0,0 +1,16 @@
+"""Generic residual layer."""
+from typing import Callable
+
+from torch import Tensor, nn
+
+
+class Residual(nn.Module):
+    """Residual layer."""
+
+    def __init__(self, fn: Callable) -> None:
+        super().__init__()
+        self.fn = fn
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies residual fn."""
+        return self.fn(x) + x
diff --git a/text_recognizer/network/transformer/__init__.py b/text_recognizer/network/transformer/__init__.py
new file mode 100644
index 0000000..a3f3011
--- /dev/null
+++ b/text_recognizer/network/transformer/__init__.py
@@ -0,0 +1 @@
+"""Transformer modules."""
diff --git a/text_recognizer/network/transformer/attend.py b/text_recognizer/network/transformer/attend.py
new file mode 100644
index 0000000..4e643fb
--- /dev/null
+++ b/text_recognizer/network/transformer/attend.py
@@ -0,0 +1,94 @@
+from typing import Optional
+from collections import namedtuple
+
+import torch
+from torch import Tensor, einsum, nn
+from einops import rearrange
+import torch.nn.functional as F
+
+Config = namedtuple(
+    "FlashAttentionConfig", ["enable_flash", "enable_math", "enable_mem_efficient"]
+)
+
+
+class Attend(nn.Module):
+    def __init__(self, use_flash: bool) -> None:
+        super().__init__()
+        self.use_flash = use_flash
+        self.cpu_cfg = Config(True, True, True)
+        self.cuda_cfg = None
+        if not torch.cuda.is_available():
+            return
+
+        device_properties = torch.cuda.get_device_properties(torch.device("cuda"))
+        if device_properties.major == 8 and device_properties.minor == 0:
+            self.cuda_cfg = Config(True, False, False)
+        else:
+            self.cuda_cfg = Config(False, True, True)
+
+    def flash_attn(self, q: Tensor, k: Tensor, v: Tensor, causal: bool) -> Tensor:
+        cfg = self.cuda_cfg if q.is_cuda else self.cpu_cfg
+        with torch.backends.cuda.sdp_kernel(**cfg._asdict()):
+            out = F.scaled_dot_product_attention(q, k, v, is_causal=causal)
+        return out
+
+    def atten(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        causal: bool,
+        mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        b = q.shape[0]
+        energy = einsum("b h i d, b h j d -> b h i j", q, k) * self.scale
+
+        mask_value = -torch.finfo(energy.dtype).max
+
+        if mask is not None:
+            energy = apply_input_mask(b, k, energy, mask, mask_value)
+
+        if causal:
+            energy = apply_causal_mask(energy, mask_value)
+
+        attn = F.softmax(energy, dim=-1)
+        attn = self.dropout(attn)
+        return einsum("b h i j, b h j d -> b h i d", attn, v)
+
+    def forward(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        causal: bool,
+        mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        if self.use_flash:
+            return self.flash_attn(q, k, v, causal)
+        else:
+            return self.atten(q, k, v, causal, mask)
+
+
+def apply_input_mask(
+    b: int,
+    k: Tensor,
+    energy: Tensor,
+    mask: Optional[Tensor],
+    mask_value: float,
+) -> Tensor:
+    """Applies an input mask."""
+    k_mask = torch.ones((b, k.shape[-2]), device=energy.device).bool()
+    q_mask = rearrange(mask, "b i -> b () i ()")
+    k_mask = rearrange(k_mask, "b j -> b () () j")
+    input_mask = q_mask * k_mask
+    return energy.masked_fill_(~input_mask, mask_value)
+
+
+def apply_causal_mask(
+    energy: Tensor,
+    mask_value: float,
+) -> Tensor:
+    """Applies a causal mask to the energy tensor."""
+    i, j, device = *energy.shape[-2:], energy.device
+    causal_mask = torch.ones((i, j), device=device, dtype=torch.bool).triu(j - i + 1)
+    return energy.masked_fill(causal_mask, mask_value)
diff --git a/text_recognizer/network/transformer/attention.py b/text_recognizer/network/transformer/attention.py
new file mode 100644
index 0000000..8e18f8a
--- /dev/null
+++ b/text_recognizer/network/transformer/attention.py
@@ -0,0 +1,56 @@
+"""Implements the attention module for the transformer."""
+from typing import Optional
+from text_recognizer.network.transformer.norm import RMSNorm
+from text_recognizer.network.transformer.attend import Attend
+
+import torch
+from einops import rearrange
+from torch import Tensor, nn
+
+
+class Attention(nn.Module):
+    """Standard attention."""
+
+    def __init__(
+        self,
+        dim: int,
+        heads: int,
+        causal: bool = False,
+        dim_head: int = 64,
+        dropout_rate: float = 0.0,
+        use_flash: bool = True,
+    ) -> None:
+        super().__init__()
+        self.heads = heads
+        inner_dim = dim_head * heads
+        self.norm = nn.LayerNorm(dim)
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(dim, inner_dim, bias=False)
+        # self.q_norm = RMSNorm(heads, dim_head)
+        # self.k_norm = RMSNorm(heads, dim_head)
+        self.attend = Attend(use_flash)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+        self.scale = dim**-0.5
+        self.causal = causal
+        self.dropout_rate = dropout_rate
+        self.dropout = nn.Dropout(p=self.dropout_rate)
+
+    def forward(
+        self,
+        x: Tensor,
+        context: Optional[Tensor] = None,
+        mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Computes the attention."""
+        x = self.norm(x)
+        q = self.to_q(x)
+        k = self.to_k(x if context is None else context)
+        v = self.to_v(x if context is None else context)
+        q, k, v = map(
+            lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), (q, k, v)
+        )
+        out = self.attend(q, k, v, self.causal, mask)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        out = self.to_out(out)
+        return out
diff --git a/text_recognizer/network/transformer/decoder.py b/text_recognizer/network/transformer/decoder.py
new file mode 100644
index 0000000..06925ba
--- /dev/null
+++ b/text_recognizer/network/transformer/decoder.py
@@ -0,0 +1,57 @@
+"""Transformer decoder module."""
+from typing import Optional
+from torch import Tensor, nn
+
+from text_recognizer.network.transformer.attention import Attention
+from text_recognizer.network.transformer.ff import FeedForward
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        inner_dim: int,
+        heads: int,
+        dim_head: int,
+        depth: int,
+        dropout_rate: float = 0.0,
+    ) -> None:
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.layers = nn.ModuleList(
+            [
+                nn.ModuleList(
+                    [
+                        Attention(
+                            dim,
+                            heads,
+                            True,
+                            dim_head,
+                            dropout_rate,
+                        ),
+                        FeedForward(dim, inner_dim, dropout_rate),
+                        Attention(
+                            dim,
+                            heads,
+                            False,
+                            dim_head,
+                            dropout_rate,
+                        ),
+                    ]
+                )
+                for _ in range(depth)
+            ]
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        context: Tensor,
+        mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        """Applies decoder block on input signals."""
+        for self_attn, ff, cross_attn in self.layers:
+            x = x + self_attn(x, mask=mask)
+            x = x + ff(x)
+            x = x + cross_attn(x, context=context)
+        return self.norm(x)
diff --git a/text_recognizer/network/transformer/embedding/__init__.py b/text_recognizer/network/transformer/embedding/__init__.py
new file mode 100644
index 0000000..bb3f904
--- /dev/null
+++ b/text_recognizer/network/transformer/embedding/__init__.py
@@ -0,0 +1 @@
+"""Positional encodings for transformers."""
diff --git a/text_recognizer/network/transformer/embedding/absolute.py b/text_recognizer/network/transformer/embedding/absolute.py
new file mode 100644
index 0000000..08b2c2a
--- /dev/null
+++ b/text_recognizer/network/transformer/embedding/absolute.py
@@ -0,0 +1,28 @@
+from typing import Optional
+
+import torch
+from torch import nn, Tensor
+from text_recognizer.network.transformer.embedding.l2_norm import l2_norm
+
+
+class AbsolutePositionalEmbedding(nn.Module):
+    def __init__(self, dim: int, max_length: int, use_l2: bool = False) -> None:
+        super().__init__()
+        self.scale = dim**-0.5 if not use_l2 else 1.0
+        self.max_length = max_length
+        self.use_l2 = use_l2
+        self.to_embedding = nn.Embedding(max_length, dim)
+        if self.use_l2:
+            nn.init.normal_(self.to_embedding.weight, std=1e-5)
+
+    def forward(self, x: Tensor, pos: Optional[Tensor] = None) -> Tensor:
+        n, device = x.shape[1], x.device
+        assert (
+            n <= self.max_length
+        ), f"Sequence length {n} is greater than the maximum positional embedding {self.max_length}"
+
+        if pos is None:
+            pos = torch.arange(n, device=device)
+
+        embedding = self.to_embedding(pos) * self.scale
+        return l2_norm(embedding) if self.use_l2 else embedding
diff --git a/text_recognizer/network/transformer/embedding/l2_norm.py b/text_recognizer/network/transformer/embedding/l2_norm.py
new file mode 100644
index 0000000..0e48bca
--- /dev/null
+++ b/text_recognizer/network/transformer/embedding/l2_norm.py
@@ -0,0 +1,9 @@
+from einops import rearrange
+import torch.nn.functional as F
+from torch import Tensor
+
+
+def l2_norm(t: Tensor, groups=1) -> Tensor:
+    t = rearrange(t, "... (g d) -> ... g d", g=groups)
+    t = F.normalize(t, p=2, dim=-1)
+    return rearrange(t, "... g d -> ... (g d)")
diff --git a/text_recognizer/network/transformer/embedding/sincos.py b/text_recognizer/network/transformer/embedding/sincos.py
new file mode 100644
index 0000000..ed6b0ab
--- /dev/null
+++ b/text_recognizer/network/transformer/embedding/sincos.py
@@ -0,0 +1,13 @@
+import torch
+
+
+def sincos_2d(h, w, dim, temperature: int = 10000, dtype=torch.float32):
+    y, x = torch.meshgrid(torch.arange(h), torch.arange(w), indexing="ij")
+    assert (dim % 4) == 0, "feature dimension must be multiple of 4 for sincos emb"
+    omega = torch.arange(dim // 4) / (dim // 4 - 1)
+    omega = 1.0 / (temperature**omega)
+
+    y = y.flatten()[:, None] * omega[None, :]
+    x = x.flatten()[:, None] * omega[None, :]
+    pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim=1)
+    return pe.type(dtype)
diff --git a/text_recognizer/network/transformer/embedding/token.py b/text_recognizer/network/transformer/embedding/token.py
new file mode 100644
index 0000000..1df2fd6
--- /dev/null
+++ b/text_recognizer/network/transformer/embedding/token.py
@@ -0,0 +1,18 @@
+from torch import nn, Tensor
+
+from text_recognizer.network.transformer.embedding.l2_norm import l2_norm
+
+
+class TokenEmbedding(nn.Module):
+    def __init__(self, num_tokens: int, dim: int, use_l2: bool = True) -> None:
+        super().__init__()
+        self.use_l2 = use_l2
+        self.to_embedding = nn.Embedding(num_tokens, dim)
+        if self.use_l2:
+            nn.init.normal_(self.to_embedding.weight, std=1e-5)
+        else:
+            nn.init.kaiming_normal_(self.to_embedding.weight)
+
+    def forward(self, x: Tensor) -> Tensor:
+        embedding = self.to_embedding(x)
+        return l2_norm(embedding) if self.use_l2 else embedding
diff --git a/text_recognizer/network/transformer/encoder.py b/text_recognizer/network/transformer/encoder.py
new file mode 100644
index 0000000..ea4b0b3
--- /dev/null
+++ b/text_recognizer/network/transformer/encoder.py
@@ -0,0 +1,46 @@
+"""Transformer encoder module."""
+from torch import Tensor, nn
+
+from text_recognizer.network.transformer.attention import Attention
+from text_recognizer.network.transformer.ff import FeedForward
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        inner_dim: int,
+        heads: int,
+        dim_head: int,
+        depth: int,
+        dropout_rate: float = 0.0,
+    ) -> None:
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.layers = nn.ModuleList(
+            [
+                nn.ModuleList(
+                    [
+                        Attention(
+                            dim,
+                            heads,
+                            False,
+                            dim_head,
+                            dropout_rate,
+                        ),
+                        FeedForward(dim, inner_dim, dropout_rate),
+                    ]
+                )
+                for _ in range(depth)
+            ]
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+    ) -> Tensor:
+        """Applies decoder block on input signals."""
+        for self_attn, ff in self.layers:
+            x = x + self_attn(x)
+            x = x + ff(x)
+        return self.norm(x)
diff --git a/text_recognizer/network/transformer/ff.py b/text_recognizer/network/transformer/ff.py
new file mode 100644
index 0000000..9181323
--- /dev/null
+++ b/text_recognizer/network/transformer/ff.py
@@ -0,0 +1,22 @@
+"""Feedforward layer in transformer."""
+from torch import Tensor, nn
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        inner_dim: int,
+        dropout_rate: float = 0.0,
+    ) -> None:
+        super().__init__()
+        self.ff = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, inner_dim),
+            nn.GELU(),
+            nn.Dropout(dropout_rate),
+            nn.Linear(inner_dim, dim),
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.ff(x)
diff --git a/text_recognizer/network/transformer/norm.py b/text_recognizer/network/transformer/norm.py
new file mode 100644
index 0000000..2737754
--- /dev/null
+++ b/text_recognizer/network/transformer/norm.py
@@ -0,0 +1,22 @@
+"""Normalization layers for transformers.
+
+Copied from lucidrains:
+    https://github.com/lucidrains/x-transformers/blob/main/x_transformers/x_transformers.py
+
+"""
+import torch
+from torch import Tensor, nn
+import torch.nn.functional as F
+
+
+class RMSNorm(nn.Module):
+    """Root mean square layer normalization."""
+
+    def __init__(self, heads: int, dim: int) -> None:
+        super().__init__()
+        self.scale = dim**-0.5
+        self.gamma = nn.Parameter(torch.ones(heads, 1, dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        """Applies normalization."""
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
diff --git a/text_recognizer/network/vit.py b/text_recognizer/network/vit.py
new file mode 100644
index 0000000..80176a8
--- /dev/null
+++ b/text_recognizer/network/vit.py
@@ -0,0 +1,76 @@
+"""Transformer module."""
+from typing import Type
+
+from einops.layers.torch import Rearrange
+from torch import Tensor, nn
+
+from text_recognizer.network.transformer.embedding.token import TokenEmbedding
+from text_recognizer.network.transformer.embedding.sincos import sincos_2d
+from text_recognizer.network.transformer.decoder import Decoder
+from text_recognizer.network.transformer.encoder import Encoder
+
+
+class VisionTransformer(nn.Module):
+    def __init__(
+        self,
+        image_height: int,
+        image_width: int,
+        patch_height: int,
+        patch_width: int,
+        dim: int,
+        num_classes: int,
+        encoder: Encoder,
+        decoder: Decoder,
+        token_embedding: TokenEmbedding,
+        pos_embedding: Type[nn.Module],
+        tie_embeddings: bool,
+        pad_index: int,
+    ) -> None:
+        super().__init__()
+        patch_dim = patch_height * patch_width
+        self.to_patch_embedding = nn.Sequential(
+            Rearrange(
+                "b c (h ph) (w pw) -> b (h w) (ph pw c)",
+                ph=patch_height,
+                pw=patch_width,
+            ),
+            nn.LayerNorm(patch_dim),
+            nn.Linear(patch_dim, dim),
+            nn.LayerNorm(dim),
+        )
+        self.patch_embedding = sincos_2d(
+            h=image_height // patch_height, w=image_width // patch_width, dim=dim
+        )
+        self.pos_embedding = pos_embedding
+        self.token_embedding = token_embedding
+        self.to_logits = (
+            nn.Linear(dim, num_classes)
+            if not tie_embeddings
+            else lambda t: t @ self.token_embedding.to_embedding.weight.t()
+        )
+        self.encoder = encoder
+        self.decoder = decoder
+        self.pad_index = pad_index
+
+    def encode(self, img: Tensor) -> Tensor:
+        x = self.to_patch_embedding(img)
+        x += self.patch_embedding.to(img.device, dtype=img.dtype)
+        return self.encoder(x)
+
+    def decode(self, text: Tensor, context: Tensor) -> Tensor:
+        text = text.long()
+        mask = text != self.pad_index
+        tokens = self.token_embedding(text)
+        tokens = tokens + self.pos_embedding(tokens)
+        output = self.decoder(tokens, context)
+        return self.to_logits(output)
+
+    def forward(
+        self,
+        img: Tensor,
+        text: Tensor,
+    ) -> Tensor:
+        """Applies decoder block on input signals."""
+        context = self.encode(img)
+        logits = self.decode(text, context)
+        return logits.permute(0, 2, 1)