/attention-augmented-conv

Implementation from the paper Attention Augmented Convolutional Networks in Tensorflow (https://arxiv.org/pdf/1904.09925v1.pdf)

Primary LanguagePython

Attention-Augmented Convolution

The tensorflow implementation from the paper Attention Augmented Convolutional Networks. Will add a Pytorch implementation of the same soon but currently the torch.einsum is very slow, even on GPU it is slower than Numpy.

Attention-Augmented Convolution

To use the layer:

from layer import augmented_conv2d

The paper has shows promising results, especially below:

ResNet-50 improvements