Keras Global Context Attention Blocks

Keras implementation of the Global Context block from the paper GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond.

Supports Conv1D, Conv2D and Conv3D directly with no modifications.

Usage

Import global_context_block from gc.py and provide it a tensor as input.

from gc import global_context_block

ip = Input(...)
x = ConvND(...)(ip)

# apply Global Context
x = global_context_block(x, reduction_ratio=16, transform_activation='linear')
...

Parameters

There are just two parameters to manage :

 - reduction_ratio: The ratio to scale the transform block.
 - transform_activation: The activation function prior to addition of the input with the context.
                         The paper uses no activation, but `sigmoid` may do better.

Requirements

  • Keras 2.2.4+
  • Tensorflow (1.13+) or CNTK

GitHub