pydrobert-gpyopt 0.0.0

# pydrobert-gpyopt
Utilities to streamline GPyOpt interfaces for ML

## How to use
GPyOpt is incredibly powerful, but a tad clunky. This lightweight package
provides two utilities in ``pydrobert.gpyopt`` to make things easier. The
first, ``GPyOptObjectiveWrapper``, wraps a function for use in GPyOpt. The
second, ``bayesopt``, takes a wrapper instance and a
``BayesianOptimizationParams`` instance and handles the optimization loop.
Here's an example:

``` python
def foo(a, d, b, **kwargs):
r = a ** d + b
weirdness = kwargs['weirdness']
if weirdness == 'flip':
r *= -1
elif weirdness == 'null':
r = 0
return r
wrapped = GPyOptObjectiveWrapper(foo)
wrapped.set_fixed_parameter('b', 1.) # 'b' will always be 1
wrapped.set_variable_parameter('a', 'continuous', (-1., 1.)) # a is real
# btw [-1,1] inc
wrapped.set_variable_parameter('d', 'discrete', (0, 3)) # d is an int
# btw [0, 3] inc
wrapped.add_parameter('weirdness') # we can add new parameters as dynamic
# keyword args if the method has a **
# parameter
wrapped.set_variable_parameter( # weirness one of the elements in the list
'weirdness', 'categorical', ('flip', 'null', None))
params = BayesianOptimizationParams(
seed=1, # setting this makes the bayesian optimization deterministic
# (assuming foo is deterministic)
log_after_iters=5,
)
best = bayesopt(wrapper, params, 'hist.csv')
```

If you provide a history file to read/write from, optimization can be
resumed after unexpected interrupts. There are a lot of options to ``bayesopt``
that are listed in ``BayesianOptimizationParams``.

## Installation

GPyOpt currently does not have a Conda build, so pydrobert-gpyopt is available
via PyPI and source install.

``` bash
pip install pydrobert-gpyopt
```