This software allows users to reproduce the results in Bridging Nonlinearities and Stochastic Regularizers with Gaussian Error Linear Units, Dan Hendrycks and Kevin Gimpel 2016.
The sigmoid(1.702 * x) * x
approximation is fast but is somewhat inaccurate. Meanwhile 0.5 * x * (1 + tanh(x * 0.7978845608 * (1 + 0.044715 * x * x)))
is slower but more accurate.
Please install Tensorflow, Lasagne, and Python 3+.