Awesome
Keras Attention Augmented Convolutions
A Keras (Tensorflow only) wrapper over the Attention Augmentation module from the paper Attention Augmented Convolutional Networks.
Provides a Layer for Attention Augmentation as well as a callable function to build a augmented convolution block.
<img src="https://github.com/titu1994/keras-attention-augmented-convs/blob/master/images/attention-augmented-convs.PNG?raw=true" height=100% width=100%>Usage
It is advisable to use the augmented_conv2d(...) function directly to build an attention augmented convolution block.
from attn_augconv import augmented_conv2d
ip = Input(...)
x = augmented_conv2d(ip, ...)
...
If you wish to add the attention module seperately, you can do so using the AttentionAugmentation1D layer as well.
from attn_augconv import AttentionAugmentation1D
ip = Input(...)
# make sure that input to the AttentionAugmentation1D layer has (2 * depth_k + depth_v) filters.
x = Conv2D(2 * depth_k + depth_v, ...)(ip)
x = AttentionAugmentation1D(depth_k, depth_v, num_heads)(x)
...
Requirements
- Tensorflow 2.0+