This is a wrapper of James Bergstra's hyperopt-convnet aimed not only at simplification but also at compliance with Nicolas Pinto's family of biologically-inspired visual features, i.e., convolutional networks (convnets).
The code has been used by our research group, but it has not been extensively tested. Therefore, one should expect some bumps while using it. We appreciate if you report them. To use the package as is, go with
python setup.py install
and refer to the scripts in simplehp/scripts to:
hp-dataset.py: run hyperparameter optimization;check-hp.py: check optimization status;protocol-eval.py: evaluate best found convnet according to the dataset's protocol.
If you want to use the package to optimize architectures on your problems of interest, then you'd better go with
python setup.py develop
and code your own data provider. This package includes an example of data provider for the PubFig83 and CalTech256 datasets in simplehp/data/rndsplits.py. You may want to use it as a starting point.
Indeed, the possibility of coding a data provider short and sweet is of great appeal in simple-hp. Have a look at simplehp/data/base.py to find this out.
Finally, in extra you will find installation scripts for installation of the whole stack of libraries required to use simple-hp. They were used in Ubuntu 12.04, and have been poorly tested. Take this as a help to get you on track. No more than that.