summaryrefslogtreecommitdiff
path: root/text_recognizer
diff options
context:
space:
mode:
authorGustaf Rydholm <gustaf.rydholm@gmail.com>2022-09-13 19:07:53 +0200
committerGustaf Rydholm <gustaf.rydholm@gmail.com>2022-09-13 19:07:53 +0200
commit9732617e402a5152ba822eb1459d6ff776aba15c (patch)
tree69cd4d4895f486bf41d20e7a74cad683ebd64294 /text_recognizer
parent3c411a90422af3e97f5b71402dcf897642b24db7 (diff)
Add convnext module
Diffstat (limited to 'text_recognizer')
-rw-r--r--text_recognizer/networks/convnext/__init__.py2
-rw-r--r--text_recognizer/networks/convnext/attention.py0
-rw-r--r--text_recognizer/networks/convnext/convnext.py75
-rw-r--r--text_recognizer/networks/convnext/downsample.py17
-rw-r--r--text_recognizer/networks/convnext/norm.py15
-rw-r--r--text_recognizer/networks/convnext/residual.py12
6 files changed, 121 insertions, 0 deletions
diff --git a/text_recognizer/networks/convnext/__init__.py b/text_recognizer/networks/convnext/__init__.py
new file mode 100644
index 0000000..8d8470f
--- /dev/null
+++ b/text_recognizer/networks/convnext/__init__.py
@@ -0,0 +1,2 @@
+"""Convnext module."""
+from text_recognizer.networks.convnext.convnext import ConvNext
diff --git a/text_recognizer/networks/convnext/attention.py b/text_recognizer/networks/convnext/attention.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/text_recognizer/networks/convnext/attention.py
diff --git a/text_recognizer/networks/convnext/convnext.py b/text_recognizer/networks/convnext/convnext.py
new file mode 100644
index 0000000..a4556a0
--- /dev/null
+++ b/text_recognizer/networks/convnext/convnext.py
@@ -0,0 +1,75 @@
+from typing import Sequence
+
+from einops import reduce, rearrange
+from einops.layers.torch import Rearrange
+import torch
+from torch import einsum, nn, Tensor
+import torch.nn.functional as F
+
+from text_recognizer.networks.convnext.downsample import Downsample
+from text_recognizer.networks.convnext.residual import Residual
+from text_recognizer.networks.convnext.norm import LayerNorm
+
+
+class ConvNextBlock(nn.Module):
+ def __init__(self, dim, dim_out, mult):
+ 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):
+ 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)),
+ ) -> None:
+ super().__init__()
+ dims = (dim, *map(lambda m: m * dim, dim_mults))
+ self.out_channels = dims[-1]
+ self.stem = nn.Conv2d(1, dims[0], kernel_size=(7, 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):
+ 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)
+ return self.norm(x)
diff --git a/text_recognizer/networks/convnext/downsample.py b/text_recognizer/networks/convnext/downsample.py
new file mode 100644
index 0000000..6e4306f
--- /dev/null
+++ b/text_recognizer/networks/convnext/downsample.py
@@ -0,0 +1,17 @@
+from typing import Tuple
+
+from einops.layers.torch import Rearrange
+from torch import nn, Tensor
+
+
+class Downsample(nn.Module):
+ 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:
+ return self.fn(x)
diff --git a/text_recognizer/networks/convnext/norm.py b/text_recognizer/networks/convnext/norm.py
new file mode 100644
index 0000000..2d896e5
--- /dev/null
+++ b/text_recognizer/networks/convnext/norm.py
@@ -0,0 +1,15 @@
+"""Layer norm for conv layers."""
+import torch
+from torch import nn, Tensor
+
+
+class LayerNorm(nn.Module):
+ def __init__(self, dim: int) -> None:
+ super().__init__()
+ self.gamma = nn.Parameter(torch.ones(1, dim, 1, 1))
+
+ def forward(self, x: Tensor) -> Tensor:
+ 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/networks/convnext/residual.py b/text_recognizer/networks/convnext/residual.py
new file mode 100644
index 0000000..3f44390
--- /dev/null
+++ b/text_recognizer/networks/convnext/residual.py
@@ -0,0 +1,12 @@
+"""Generic residual layer."""
+from typing import Callable
+from torch import nn, Tensor
+
+
+class Residual(nn.Module):
+ def __init__(self, fn: Callable) -> None:
+ super().__init__()
+ self.fn = fn
+
+ def forward(self, x: Tensor) -> Tensor:
+ return self.fn(x) + x