Implementation of Single Headed Attention - Recurrent Neural Networks in Julia and Knet.

Stephan Merity. Single Headed Attention RNN: Stop Thinking With Your Head. arXiv preprint arXiv:1911.11423, 2019.

https://arxiv.org/abs/1911.11423v2

After downloading the data and preprocessing it using

sh getdata.sh

You can train the main model of SHA-RNN paper by either:

running sharnn-main.jl in shell

cd examples
julia sharnn-main.jl

or using SHA-RNN notebook.

This implementation is identical to the one of Smerity's original implementation sha-rnn.

But it is slower, since it does not use the same performance tricks that the version of SHA-RNN that was implemented using pytorch uses.

• Fused layer normalization (check if Apex CUDA code can be used with Knet)
• Using half precision floating point (Float16) for memory efficiency
• Checkpoint feature similar to pytorch's checkpoint.