Basics
"""
Code to simplify sampling the Gaussian process.
"""
import numpy as np
[docs]def draw_samples(gp, **kwargs):
"""
Construct an array to pass to the Gaussian process to pull out a number of samples from a
high dimensional GP.
Parameters
----------
gp : GPR object
The Gaussian process object
kwargs : int, list, or tuple
The ranges for each value.
In the format
parameter = 0.9
the axis will be constant
Alternatively a list or tuple can be passed to form a range (which uses the same
arrangement of arguments as numpy's linspace function:
parameter = (0.0, 1.5, 100)
will produce an axis 100 samples long.
"""
parameters = set(kwargs.keys())
gp_parameters = set(gp.training_object.target_names)
if len(parameters.difference(gp_parameters))>0:
raise UndefinedParameters
column_values = {}
dimensions = 0
len_vec = 1
rangecols = []
for parameter in kwargs.items():
name = parameter[0]
value = parameter[1]
column = gp.training_object.name2ix(name)
# make all of the axes for the sampling
if isinstance(value, float) or isinstance(value, int):
# The passed value is a single number, so it should be repeated for every
# sample
column_values[column] = value
#elif "__iter__" in dir(value):
else:
# The passed value is something like a list, so we'll use it to construct an
# axis using np.linspace
if len_vec > 1 and dimensions >= 2:
raise UnsupportedError
if len(value) == 2:
column_values[column] = np.linspace(value[0], value[1])
elif len(value) == 3:
column_values[column] = np.linspace(value[0], value[1], value[2])
rangecols.append(column)
len_vec = len(column_values[column])
dimensions += 1
if dimensions <= 1:
output = np.zeros( (len_vec, len(kwargs.items())))
elif dimensions == 2:
gridpoints = np.meshgrid(column_values[rangecols[0]], column_values[rangecols[1]])
points = np.dstack(gridpoints).reshape(-1, 2)
column_values[rangecols[0]] = points[:,0]
column_values[rangecols[1]] = points[:,1]
output = np.zeros( (len(points), len(kwargs.items())))
for column in column_values.items():
output[:, column[0]] = column[1]
return output