Reproduce sparse-evolutionary-artificial-neural-networks

[drplanar]

• Forked https://github.com/liuchangjohn/sparse-evolutionary-artificial-neural-networks
• Install keras-contrib from github
• Run sparse-evolutionary-artificial-neural-networks/SET-MLP-Keras-Weights-Mask/dense_mlp_keras_cifar10.py. It uses keras to download cifar10 data automatically. This is the dense MLP baseline.

Dense baseline took ~5 hours to run. Result is quite close to that reported in the paper.

[drplanar]

The main file https://github.com/liuchangjohn/sparse-evolutionary-artificial-neural-networks/blob/master/SET-MLP-Keras-Weights-Mask/dense_mlp_keras_cifar10.py is very short. Worth taking a look.

[drplanar]

Ran set_mlp_srelu_sgd_cifar10_acc.txt and reproduced the sparse result (72.4% vs 70% dense). Took about 6 hours.