Working on new implementation of efficientnet

3 files changed, 185 insertions, 21 deletions

diff --git a/text_recognizer/networks/encoders/efficientnet/efficientnet.py b/text_recognizer/networks/encoders/efficientnet/efficientnet.py
new file mode 100644
index 0000000..d953c10
--- /dev/null
+++ b/text_recognizer/networks/encoders/efficientnet/efficientnet.py
@@ -0,0 +1,9 @@
+"""Efficient net."""
+from torch import nn, Tensor
+
+
+class EfficientNet(nn.Module):
+    def __init__(
+        self,
+    ) -> None:
+        super().__init__()
diff --git a/text_recognizer/networks/encoders/efficientnet/mbconv_block.py b/text_recognizer/networks/encoders/efficientnet/mbconv.py
index c501777..602aeb7 100644
--- a/text_recognizer/networks/encoders/efficientnet/mbconv_block.py
+++ b/text_recognizer/networks/encoders/efficientnet/mbconv.py
@@ -5,7 +5,7 @@ import torch
 from torch import nn, Tensor
 from torch.nn import functional as F
 
-from .utils import get_same_padding_conv2d
+from .utils import calculate_output_image_size, drop_connection, get_same_padding_conv2d
 
 
 class MBConvBlock(nn.Module):
@@ -22,19 +22,23 @@ class MBConvBlock(nn.Module):
         se_ratio: float,
         id_skip: bool,
         expand_ratio: int,
-        image_size: Tuple[int, int],
+        image_size: Optional[Tuple[int, int]],
     ) -> None:
         super().__init__()
         self.kernel_size = kernel_size
         self.bn_momentum = bn_momentum
         self.bn_eps = bn_eps
         self.id_skip = id_skip
-        (
-            self._inverted_bottleneck,
-            self._depthwise,
-            self._squeeze_excite,
-            self._pointwise,
-        ) = self._build(
+        self.in_channels = self.in_channels
+        self.out_channels = out_channels
+
+        # Placeholders for layers.
+        self._inverted_bottleneck: nn.Sequential = None
+        self._depthwise: nn.Sequential = None
+        self._squeeze_excite: nn.Sequential = None
+        self._pointwise: nn.Sequential = None
+
+        self._build(
             image_size=image_size,
             in_channels=in_channels,
             out_channels=out_channels,
@@ -46,19 +50,17 @@ class MBConvBlock(nn.Module):
 
     def _build(
         self,
-        image_size: Tuple[int, int],
+        image_size: Optional[Tuple[int, int]],
         in_channels: int,
         out_channels: int,
         kernel_size: int,
         stride: int,
         expand_ratio: int,
         se_ratio: float,
-    ) -> Tuple[
-        Optional[nn.Sequential], nn.Sequential, Optional[nn.Sequential], nn.Sequential
-    ]:
+    ) -> None:
         has_se = se_ratio is not None and 0.0 < se_ratio < 1.0
         inner_channels = in_channels * expand_ratio
-        inverted_bottleneck = (
+        self._inverted_bottleneck = (
             self._configure_inverted_bottleneck(
                 image_size=image_size,
                 in_channels=in_channels,
@@ -68,7 +70,7 @@ class MBConvBlock(nn.Module):
             else None
         )
 
-        depthwise = self._configure_depthwise(
+        self._depthwise = self._configure_depthwise(
             image_size=image_size,
             in_channels=inner_channels,
             out_channels=inner_channels,
@@ -78,7 +80,7 @@ class MBConvBlock(nn.Module):
         )
 
         image_size = calculate_output_image_size(image_size, stride)
-        squeeze_excite = (
+        self._squeeze_excite = (
             self._configure_squeeze_excite(
                 in_channels=inner_channels,
                 out_channels=inner_channels,
@@ -88,14 +90,13 @@ class MBConvBlock(nn.Module):
             else None
         )
 
-        pointwise = self._configure_pointwise(
+        self._pointwise = self._configure_pointwise(
             image_size=image_size, in_channels=inner_channels, out_channels=out_channels
         )
-        return inverted_bottleneck, depthwise, squeeze_excite, pointwise
 
     def _configure_inverted_bottleneck(
         self,
-        image_size: Tuple[int, int],
+        image_size: Optional[Tuple[int, int]],
         in_channels: int,
         out_channels: int,
     ) -> nn.Sequential:
@@ -116,7 +117,7 @@ class MBConvBlock(nn.Module):
 
     def _configure_depthwise(
         self,
-        image_size: Tuple[int, int],
+        image_size: Optional[Tuple[int, int]],
         in_channels: int,
         out_channels: int,
         groups: int,
@@ -159,7 +160,7 @@ class MBConvBlock(nn.Module):
         )
 
     def _configure_pointwise(
-        self, image_size: Tuple[int, int], in_channels: int, out_channels: int
+        self, image_size: Optional[Tuple[int, int]], in_channels: int, out_channels: int
     ) -> nn.Sequential:
         Conv2d = get_same_padding_conv2d(image_size=image_size)
         return nn.Sequential(
@@ -174,7 +175,18 @@ class MBConvBlock(nn.Module):
             ),
         )
 
-    def forward(self, x: Tensor, drop_connection_rate: Optional[float]) -> Tensor:
+    def _stochastic_depth(
+        self, x: Tensor, residual: Tensor, drop_connection_rate: Optional[float]
+    ) -> Tensor:
+        if self.id_skip and self.stride == 1 and self.in_channels == self.out_channels:
+            if drop_connection_rate:
+                x = drop_connection(x, p=drop_connection_rate, training=self.training)
+            x += residual
+        return x
+
+    def forward(
+        self, x: Tensor, drop_connection_rate: Optional[float] = None
+    ) -> Tensor:
         residual = x
         if self._inverted_bottleneck is not None:
             x = self._inverted_bottleneck(x)
@@ -189,3 +201,5 @@ class MBConvBlock(nn.Module):
         x = self._pointwise(x)
 
         # Stochastic depth
+        x = self._stochastic_depth(x, residual, drop_connection_rate)
+        return x
diff --git a/text_recognizer/networks/encoders/efficientnet/utils.py b/text_recognizer/networks/encoders/efficientnet/utils.py
new file mode 100644
index 0000000..4b4a787
--- /dev/null
+++ b/text_recognizer/networks/encoders/efficientnet/utils.py
@@ -0,0 +1,141 @@
+"""Util functions for efficient net."""
+from functools import partial
+import math
+from typing import Any, Optional, Tuple, Type
+
+import torch
+from torch import nn, Tensor
+import torch.functional as F
+
+
+def calculate_output_image_size(
+    image_size: Optional[Tuple[int, int]], stride: int
+) -> Optional[Tuple[int, int]]:
+    """Calculates the output image size when using conv2d with same padding."""
+    if image_size is None:
+        return None
+    height = int(math.ceil(image_size[0] / stride))
+    width = int(math.ceil(image_size[1] / stride))
+    return height, width
+
+
+def drop_connection(x: Tensor, p: float, training: bool) -> Tensor:
+    """Drop connection.
+
+    Drops the entire convolution with a given survival probability.
+
+    Args:
+        x (Tensor): Input tensor.
+        p (float): Survival probability between 0.0 and 1.0.
+        training (bool): The running mode.
+
+    Shapes:
+        - x: :math: `(B, C, W, H)`.
+        - out: :math: `(B, C, W, H)`.
+
+        where B is the batch size, C is the number of channels, W is the width, and H
+        is the height.
+
+    Returns:
+        out (Tensor): Output after drop connection.
+    """
+    assert 0.0 <= p <= 1.0, "p must be in range of [0, 1]"
+
+    if not training:
+        return x
+
+    bsz = x.shape[0]
+    survival_prob = 1 - p
+
+    # Generate a binary tensor mask according to probability (p for 0, 1-p for 1)
+    random_tensor = survival_prob
+    random_tensor += torch.rand([bsz, 1, 1, 1]).type_as(x)
+    binary_tensor = torch.floor(random_tensor)
+
+    out = x / survival_prob * binary_tensor
+    return out
+
+
+def get_same_padding_conv2d(image_size: Optional[Tuple[int, int]]) -> Type[nn.Conv2d]:
+    if image_size is None:
+        return Conv2dDynamicSamePadding
+    return partial(Conv2dStaticSamePadding, image_size=image_size)
+
+
+class Conv2dDynamicSamePadding(nn.Conv2d):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = 1,
+        dilation: int = 1,
+        groups: int = 1,
+        bias: bool = True,
+    ) -> None:
+        super().__init__(
+            in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias
+        )
+        self.stride = [self.stride] * 2
+
+    def forward(self, x: Tensor) -> Tensor:
+        ih, iw = x.shape[-2:]
+        kh, kw = self.weight.shape[-2:]
+        sh, sw = self.stride
+        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)
+        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
+        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
+        if pad_h > 0 or pad_w > 0:
+            x = F.pad(
+                x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2]
+            )
+        return F.conv2d(
+            x,
+            self.weight,
+            self.bias,
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+
+class Conv2dStaticSamePadding(nn.Conv2d):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        image_size: Tuple[int, int],
+        stride: int = 1,
+        **kwargs: Any
+    ):
+        super().__init__(in_channels, out_channels, kernel_size, stride, **kwargs)
+        self.stride = [self.stride] * 2
+
+        # Calculate padding based on image size and save it.
+        ih, iw = image_size
+        kh, kw = self.weight.shape[-2:]
+        sh, sw = self.stride
+        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)
+        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
+        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
+        if pad_h > 0 or pad_w > 0:
+            self.static_padding = nn.ZeroPad2d(
+                (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2)
+            )
+        else:
+            self.static_padding = nn.Identity()
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.static_padding(x)
+        x = F.pad(
+            x,
+            self.weight,
+            self.bias,
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+        return x