c3/libraries/sampling.py
"""Functions to select samples from a dataset by various criteria."""
import random
import numpy as np
sampling = dict()
def sampling_reg_deco(func):
"""
Decorator for making registry of functions
"""
sampling[str(func.__name__)] = func
return func
@sampling_reg_deco
def all(learn_from, batch_size):
"""
Return all points.
Parameters
----------
learn_from : list
List of data points
batch_size : int
Number of points to select
Returns
-------
list
All indeces.
"""
total_size = len(learn_from)
all = list(range(total_size))
return all
@sampling_reg_deco
def from_start(learn_from, batch_size):
"""
Select from the beginning.
Parameters
----------
learn_from : list
List of data points
batch_size : int
Number of points to select
Returns
-------
list
Selected indices.
"""
total_size = len(learn_from)
all = list(range(total_size))
return all[:batch_size]
@sampling_reg_deco
def from_end(learn_from, batch_size):
"""
Select from the end.
Parameters
----------
learn_from : list
List of data points
batch_size : int
Number of points to select
Returns
-------
list
Selected indices.
"""
total_size = len(learn_from)
all = list(range(total_size))
return all[-batch_size:]
@sampling_reg_deco
def even(learn_from, batch_size):
"""
Select evenly distanced samples across the set.
Parameters
----------
learn_from : list
List of data points
batch_size : int
Number of points to select
Returns
-------
list
Selected indices.
"""
total_size = len(learn_from)
all = list(range(total_size))
n = int(np.ceil(total_size / batch_size))
return all[::n]
@sampling_reg_deco
def random_sample(learn_from, batch_size):
"""
Select randomly.
Parameters
----------
learn_from : list
List of data points.
batch_size : int
Number of points to select.
Returns
-------
list
Selected indices.
"""
total_size = len(learn_from)
all = list(range(total_size))
return random.sample(all, batch_size)
@sampling_reg_deco
def high_std(learn_from, batch_size):
"""
Select points that have a high ratio of standard deviation to mean. Sampling from ORBIT data, points with a high
std have the most coherent error, thus might be suitable for model learning. This has yet to be confirmed beyond
anecdotal observation.
Parameters
----------
learn_from : list
List of data points.
batch_size : int
Number of points to select.
Returns
-------
list
Selected indices.
"""
a_val = []
for sample in learn_from:
a_val.append(np.std(sample["results"]) / np.mean(sample["results"]))
return np.argsort(np.array(a_val))[-batch_size:]
@sampling_reg_deco
def even_fid(learn_from, batch_size):
"""
Select evenly among reached fidelities.
Parameters
----------
learn_from : list
List of data points.
batch_size : int
Number of points to select.
Returns
-------
list
Selected indices.
"""
total_size = len(learn_from)
res = []
for sample in learn_from:
res.append(np.mean(sample["results"]))
n = int(np.ceil(total_size / batch_size))
return np.argsort(np.array(res))[::n]