doctaphred/umami

Bayesian optimization of machine learning models.

Umami

Umami is a package to perform Bayesian optimization according to the algorithms outlined in the paper:

Practical Bayesian Optimization of Machine Learning Algorithms

Jasper Snoek, Hugo Larochelle and Ryan P. Adams

Advances in Neural Information Processing Systems, 2012

Umami is a fork of the GPLv3+ version of Spearmint. I named it Umami because I am bad at naming projects.

I chose to fork Spearmint because development moved to an academic-use-only license and the project seems to have died.

Goals

• Port to Python 3 (only)
• Clean up the source.
• Prune code to a maintainable subset.
• Improve the user experience.

I have gotten Umami into a state where it actually runs (and optimizes!) local Python tasks, but there is a lot of work left. I plan on porting to Python 3. Right now it seems the biggest blocker is Weave which is used for a single for loop which could be written with better usage Numpy or maybe Numba if I can't figure that out.

While cleaning up the source code I plan to prune the features down to something more manageable. I plan on only supporting local execution of Python files. If you want to run a task in a different language you could write a little Python file to wrap the calls to your language, Umami doesn't need to support this directly.

I would like to improve the output of Umami. Currently it just dumps a bunch of files in the experiment directory but this could be stored in a standard location or format.

Usage

Umami uses a config.json file which is mostly like Spearmint. Currently Umami only supports Python and local execution so these options are ignored in the config.json file.

Umami can be run on the command line with:

$ python -m umami /path/to/config.json [OPTIONS]

Example

Below is an example config.json which show different variable types and shapes.

{
    "main-file"       : "run_task.py",
    "name"            : "mlp-regressor-all-solvers",
    "likelihood"      : "GAUSSIAN",
    "variables" : {
        "alpha" : {
            "type" : "FLOAT",
            "size" : 1,
            "min"  : 0.00001,
            "max"  : 0.009
        },
        "hidden_layer_sizes" : {
            "type" : "INT",
            "size" : 3,
            "min"  : 1,
            "max"  : 200
        },
        "solver": {
            "type": "ENUM",
            "size": 1,
            "options": ["lbfgs", "sgd", "adam"]
        },
        "activation": {
            "type": "ENUM",
            "size": 1,
            "options": ["identity", "logistic", "tanh", "relu"]
        }
    }
}

The main-file must define a function main like:

def main(job_id, params):
    # train
    return error

params is passed as a dictionary from variable name to values based on the variable spaces defined in config.json.