Add 'Efficient Channel Attention'

README.md

@@ -48,6 +48,7 @@
 - [x] `Gather-Excite` - [Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks](https://arxiv.org/abs/1810.12348), NeurIPS, 2018
 - [x] `CBAM` - [CBAM: Convolutional Block Attention Module](https://arxiv.org/abs/1807.06521), ECCV, 2018
 - [x] `SKNets` - [Selective Kernel Networks](https://arxiv.org/abs/1903.06586), CVPR, 2019
+- [x] `ECA` - [ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks](https://arxiv.org/abs/1910.03151), CVPR, 2019
 
 ### Transformer
 

cvm/models/ops/blocks/__init__.py

@@ -18,4 +18,5 @@
 from .interpolate import Interpolate
 from .gather_excite import GatherExciteBlock
 from .selective_kernel import SelectiveKernelBlock
-from .cbam import CBAM
+from .cbam import CBAM
+from .efficient_channel_attention import EfficientChannelAttention

cvm/models/ops/blocks/efficient_channel_attention.py

@@ -0,0 +1,31 @@
+import math
+import torch
+from torch import nn
+
+
+class EfficientChannelAttention(nn.Module):
+    r"""
+    Paper: ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks, https://arxiv.org/abs/1910.03151
+    """
+    def __init__(
+        self,
+        in_channels,
+        gamma=2,
+        beta=2
+    ) -> None:
+        super().__init__()
+
+        t = int(abs((math.log(in_channels, 2) + beta) / gamma))
+        k = max(t if t % 2 else t + 1, 3)
+
+        self.pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.conv = nn.Conv1d(1, 1, kernel_size=k, padding=(k - 1) // 2)
+        self.gate = nn.Sigmoid()
+
+    def forward(self, x: torch.Tensor):
+        y = self.pool(x)
+        y = self.conv(y.view(y.shape[0], 1, -1))
+        y = y.view(y.shape[0], -1, 1, 1)
+        y = self.gate(y)
+
+        return x * y.expand_as(x)