Easily write and train Maxout networks
This project implements: Ian Goodfellow et al. “Maxout Networks”, [http://proceedings.mlr.press/v28/goodfellow13.html]
Make sure you have Tensoflow (version 1.13, at least) and Numpy installed. Run:
python3 maxout.py -h
to see all options and commands available.
The input of this program is a dataset in the "datasets/" directory. The datasets now supported are MNIST and CIFAR-10. To use them, download their binary version to "datasets/MNIST" and "datasets/CIFAR-10" directories. You should also create a validation split and set the correct filenames. To use different datasets, write a loader in "data.py" that follows the same Tensorflow data API.
The outputs are Tensorboard logs in "logs/" and model parameters in "models/" directories. You can customize the nets or write your own in "nets/*_net.py" files.