nits

.gitignore

@@ -1,2 +1,3 @@
 __pycache__
-.pytest_cache
+.pytest_cache
+dist/

README.md

@@ -0,0 +1,18 @@
+# k3-addons: Additional multi-backend functionality for Keras 3.
+![Logo](.assets/k3-addons.png)
+
+# Includes:
+- Layers
+ - Pooling:
+ - `AdaptiveAveragePooling1D`
+ - `AdaptiveMaxPooling1D`
+ - `AdaptiveAveragePooling2D`
+ - `AdaptiveMaxPooling2D`
+ - Attention:
+ - `DoubleAttention`
+ - `AFTFull`
+ - `ChannelAttention2D`
+ - `SpatialAttention2D`
+ - `ECAAttention`
+ - `ExternalAttention`
+ - `ResidualAttention`

k3_addons/layers/attention/cbam.py

@@ -0,0 +1,68 @@
+from keras import layers, ops, Sequential, backend
+
+from k3_addons.layers.pooling.adaptive_pooling import (
+ AdaptiveMaxPool2D,
+ AdaptiveAveragePool2D,
+)
+from k3_addons.api_export import k3_export
+
+@k3_export('k3_addons.layers.ChannelAttention2D')
+class ChannelAttention(layers.Layer):
+ def __init__(self, reduction=16):
+ super().__init__()
+ self.reduction = reduction
+ self.maxpool = AdaptiveMaxPool2D(1)
+ self.avgpool = AdaptiveAveragePool2D(1)
+ self.data_format = backend.image_data_format()
+
+ def build(self, input_shape):
+ if self.data_format == "channels_last":
+ input_dim = input_shape[3]
+ else:
+ input_dim = input_shape[1]
+ self.se = Sequential(
+ [
+ layers.Conv2D(input_dim // self.reduction, 1, use_bias=False),
+ layers.Activation("relu"),
+ layers.Conv2D(input_dim, 1, use_bias=False),
+ ]
+ )
+
+ def call(self, x):
+ max_result = self.maxpool(x)
+ avg_result = self.avgpool(x)
+ max_out = self.se(max_result)
+ avg_out = self.se(avg_result)
+ output = ops.sigmoid(max_out + avg_out)
+ return output
+
+@k3_export('k3_addons.layers.SpatialAttention2D')
+class SpatialAttention(layers.Layer):
+ def __init__(self, kernel_size=7):
+ super().__init__()
+ self.conv = layers.Conv2D(1, kernel_size=kernel_size, padding='same')
+
+ def call(self, x):
+ if backend.image_data_format() == "channels_first":
+ axis = 1
+ else:
+ axis = 3
+ max_result = ops.max(x, axis=axis, keepdims=True)
+ avg_result = ops.mean(x, axis=axis, keepdims=True)
+ result = ops.concatenate([max_result, avg_result], axis)
+ output = self.conv(result)
+ output = ops.sigmoid(output)
+ return output
+
+@k3_export('k3_addons.layers.CBAM')
+class CBAMBlock(layers.Layer):
+ def __init__(self, reduction=16, kernel_size=49):
+ super().__init__()
+ self.channel_attention = ChannelAttention(reduction=reduction)
+ self.spatial_attention = SpatialAttention(kernel_size=kernel_size)
+
+ def call(self, x):
+ residual = x
+ out = x * self.channel_attention(x)
+ out = out * self.spatial_attention(out)
+ return out + residual

k3_addons/layers/attention/eca.py

@@ -0,0 +1,23 @@
+from keras import layers, ops, Sequential
+from k3_addons.layers.pooling.adaptive_pooling import AdaptiveAveragePool2D
+from k3_addons.api_export import k3_export
+
+
+@k3_export('k3_addons.layers.ECAAttention')
+class ECAAttention(layers.Layer):
+
+ def __init__(self, kernel_size=3):
+ super().__init__()
+ self.pooling = AdaptiveAveragePool2D(1)
+ self.conv=layers.Conv1D(1, kernel_size=kernel_size, padding='same')
+
+
+ def call(self, x):
+ y = self.pooling(x) #b,1, 1, c
+ y = ops.squeeze(y, axis=2) # b, 1, c
+ y = ops.transpose(y, axes=[0, 2, 1]) # b, c, 1
+ y = self.conv(y) #b, c, 1
+ y = ops.sigmoid(y) # b, c, 1
+ y = ops.transpose(y, axes=[0, 2, 1]) # b, 1, c
+ y= ops.expand_dims(y, axis=2) #bs, 1, 1, c
+ return x * ops.broadcast_to(y, ops.shape(x))