Revert "Delete convnext"

This reverts commit 7239bce214607c70a7a91358586f265b2f74de7b.

7 files changed, 466 insertions, 0 deletions

diff --git a/notebooks/04-convnext.ipynb b/notebooks/04-convnext.ipynb
new file mode 100644
index 0000000..5ab71c8
--- /dev/null
+++ b/notebooks/04-convnext.ipynb
@@ -0,0 +1,248 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "id": "7c02ae76-b540-4b16-9492-e9210b3b9249",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The autoreload extension is already loaded. To reload it, use:\n",
+      "  %reload_ext autoreload\n"
+     ]
+    }
+   ],
+   "source": [
+    "import os\n",
+    "os.environ['CUDA_VISIBLE_DEVICE'] = ''\n",
+    "import random\n",
+    "\n",
+    "%matplotlib inline\n",
+    "import matplotlib.pyplot as plt\n",
+    "\n",
+    "import numpy as np\n",
+    "from omegaconf import OmegaConf\n",
+    "\n",
+    "%load_ext autoreload\n",
+    "%autoreload 2\n",
+    "\n",
+    "from importlib.util import find_spec\n",
+    "if find_spec(\"text_recognizer\") is None:\n",
+    "    import sys\n",
+    "    sys.path.append('..')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 15,
+   "id": "ccdb6dde-47e5-429a-88f2-0764fb7e259a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from hydra import compose, initialize\n",
+    "from omegaconf import OmegaConf\n",
+    "from hydra.utils import instantiate"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "id": "3cf50475-39f2-4642-a7d1-5bcbc0a036f7",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "path = \"../training/conf/network/convnext.yaml\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 37,
+   "id": "e52ecb01-c975-4e55-925d-1182c7aea473",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "with open(path, \"rb\") as f:\n",
+    "    cfg = OmegaConf.load(f)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 38,
+   "id": "f939aa37-7b1d-45cc-885c-323c4540bda1",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "{'_target_': 'text_recognizer.network.convnext.ConvNext', 'dim': 16, 'dim_mults': [2, 4, 8], 'depths': [3, 3, 6], 'downsampling_factors': [[2, 2], [2, 2], [2, 2]], 'attn': {'_target_': 'text_recognizer.network.convnext.TransformerBlock', 'attn': {'_target_': 'text_recognizer.network.convnext.Attention', 'dim': 128, 'heads': 4, 'dim_head': 64, 'scale': 8}, 'ff': {'_target_': 'text_recognizer.network.convnext.FeedForward', 'dim': 128, 'mult': 4}}}"
+      ]
+     },
+     "execution_count": 38,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "cfg"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 19,
+   "id": "a2b420c1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "cfg.dim_mults = [2, 4, 8, 8]\n",
+    "cfg.depths = [3, 3, 6, 6]\n",
+    "cfg.downsampling_factors = [[2, 2], [2, 2], [2, 2], [2, 1]]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 39,
+   "id": "c9589350",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "net = instantiate(cfg)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 40,
+   "id": "618b997c-e6a6-4487-b70c-9d260cb556d3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from torchinfo import summary"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 41,
+   "id": "25759b7b-8deb-4163-b75d-a1357c9fe88f",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "====================================================================================================\n",
+       "Layer (type:depth-idx)                             Output Shape              Param #\n",
+       "====================================================================================================\n",
+       "ConvNext                                           [2, 128, 72, 80]          165,408\n",
+       "├─TransformerBlock: 1-5                            [2, 128, 72, 80]          (recursive)\n",
+       "│    └─Attention: 2-6                              [2, 128, 72, 80]          (recursive)\n",
+       "│    │    └─LayerNorm: 3-13                        [2, 128, 72, 80]          (recursive)\n",
+       "├─Conv2d: 1-3                                      [2, 16, 576, 640]         (recursive)\n",
+       "├─TransformerBlock: 1                              --                        --\n",
+       "│    └─Attention: 2                                --                        --\n",
+       "│    │    └─Conv2d: 3-15                           [2, 128, 72, 80]          (recursive)\n",
+       "│    └─FeedForward: 2-7                            [2, 128, 72, 80]          (recursive)\n",
+       "│    │    └─Residual: 3-16                         [2, 128, 72, 80]          (recursive)\n",
+       "│    │    │    └─Sequential: 4-26                  [2, 128, 72, 80]          (recursive)\n",
+       "├─Conv2d: 1-3                                      [2, 16, 576, 640]         (recursive)\n",
+       "├─ModuleList: 1-4                                  --                        --\n",
+       "│    └─ModuleList: 2-3                             --                        --\n",
+       "│    │    └─ConvNextBlock: 3-4                     [2, 16, 576, 640]         --\n",
+       "│    │    │    └─Conv2d: 4-2                       [2, 16, 576, 640]         800\n",
+       "│    │    │    └─Sequential: 4-3                   [2, 16, 576, 640]         9,280\n",
+       "│    │    │    └─Identity: 4-4                     [2, 16, 576, 640]         --\n",
+       "│    │    └─ModuleList: 3-5                        --                        --\n",
+       "│    │    │    └─ConvNextBlock: 4-5                [2, 16, 576, 640]         10,080\n",
+       "│    │    │    └─ConvNextBlock: 4-6                [2, 16, 576, 640]         10,080\n",
+       "│    │    │    └─ConvNextBlock: 4-7                [2, 16, 576, 640]         10,080\n",
+       "│    │    └─Downsample: 3-6                        [2, 32, 288, 320]         --\n",
+       "│    │    │    └─Sequential: 4-8                   [2, 32, 288, 320]         2,080\n",
+       "│    └─ModuleList: 2-4                             --                        --\n",
+       "│    │    └─ConvNextBlock: 3-7                     [2, 32, 288, 320]         --\n",
+       "│    │    │    └─Conv2d: 4-9                       [2, 32, 288, 320]         1,600\n",
+       "│    │    │    └─Sequential: 4-10                  [2, 32, 288, 320]         36,992\n",
+       "│    │    │    └─Identity: 4-11                    [2, 32, 288, 320]         --\n",
+       "│    │    └─ModuleList: 3-8                        --                        --\n",
+       "│    │    │    └─ConvNextBlock: 4-12               [2, 32, 288, 320]         38,592\n",
+       "│    │    │    └─ConvNextBlock: 4-13               [2, 32, 288, 320]         38,592\n",
+       "│    │    │    └─ConvNextBlock: 4-14               [2, 32, 288, 320]         38,592\n",
+       "│    │    └─Downsample: 3-9                        [2, 64, 144, 160]         --\n",
+       "│    │    │    └─Sequential: 4-15                  [2, 64, 144, 160]         8,256\n",
+       "│    └─ModuleList: 2-5                             --                        --\n",
+       "│    │    └─ConvNextBlock: 3-10                    [2, 64, 144, 160]         --\n",
+       "│    │    │    └─Conv2d: 4-16                      [2, 64, 144, 160]         3,200\n",
+       "│    │    │    └─Sequential: 4-17                  [2, 64, 144, 160]         147,712\n",
+       "│    │    │    └─Identity: 4-18                    [2, 64, 144, 160]         --\n",
+       "│    │    └─ModuleList: 3-11                       --                        --\n",
+       "│    │    │    └─ConvNextBlock: 4-19               [2, 64, 144, 160]         150,912\n",
+       "│    │    │    └─ConvNextBlock: 4-20               [2, 64, 144, 160]         150,912\n",
+       "│    │    │    └─ConvNextBlock: 4-21               [2, 64, 144, 160]         150,912\n",
+       "│    │    │    └─ConvNextBlock: 4-22               [2, 64, 144, 160]         150,912\n",
+       "│    │    │    └─ConvNextBlock: 4-23               [2, 64, 144, 160]         150,912\n",
+       "│    │    │    └─ConvNextBlock: 4-24               [2, 64, 144, 160]         150,912\n",
+       "│    │    └─Downsample: 3-12                       [2, 128, 72, 80]          --\n",
+       "│    │    │    └─Sequential: 4-25                  [2, 128, 72, 80]          32,896\n",
+       "├─TransformerBlock: 1-5                            [2, 128, 72, 80]          (recursive)\n",
+       "│    └─Attention: 2-6                              [2, 128, 72, 80]          (recursive)\n",
+       "│    │    └─LayerNorm: 3-13                        [2, 128, 72, 80]          (recursive)\n",
+       "│    │    └─Conv2d: 3-14                           [2, 768, 72, 80]          98,304\n",
+       "│    │    └─Conv2d: 3-15                           [2, 128, 72, 80]          (recursive)\n",
+       "│    └─FeedForward: 2-7                            [2, 128, 72, 80]          (recursive)\n",
+       "│    │    └─Residual: 3-16                         [2, 128, 72, 80]          (recursive)\n",
+       "│    │    │    └─Sequential: 4-26                  [2, 128, 72, 80]          (recursive)\n",
+       "├─LayerNorm: 1-6                                   [2, 128, 72, 80]          128\n",
+       "====================================================================================================\n",
+       "Total params: 1,558,144\n",
+       "Trainable params: 1,558,144\n",
+       "Non-trainable params: 0\n",
+       "Total mult-adds (G): 114.00\n",
+       "====================================================================================================\n",
+       "Input size (MB): 2.95\n",
+       "Forward/backward pass size (MB): 3822.06\n",
+       "Params size (MB): 5.57\n",
+       "Estimated Total Size (MB): 3830.58\n",
+       "===================================================================================================="
+      ]
+     },
+     "execution_count": 41,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "summary(net, (2, 1, 576, 640), device=\"cpu\", depth=4)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "05c1d499",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.4"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}
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