import itertools
from collections import defaultdict
from collections.abc import Iterable
from contextlib import suppress
from functools import partial
from operator import itemgetter
import numpy as np
from adaptive.learner.base_learner import BaseLearner
from adaptive.notebook_integration import ensure_holoviews
from adaptive.utils import cache_latest, named_product, restore
def dispatch(child_functions, arg):
index, x = arg
return child_functions[index](x)
class BalancingLearner(BaseLearner):
r"""Choose the optimal points from a set of learners.
Parameters
----------
learners : sequence of `~adaptive.BaseLearner`\s
The learners from which to choose. These must all have the same type.
cdims : sequence of dicts, or (keys, iterable of values), optional
Constant dimensions; the parameters that label the learners. Used
in `plot`.
Example inputs that all give identical results:
- sequence of dicts:
>>> cdims = [{'A': True, 'B': 0},
... {'A': True, 'B': 1},
... {'A': False, 'B': 0},
... {'A': False, 'B': 1}]`
- tuple with (keys, iterable of values):
>>> cdims = (['A', 'B'], itertools.product([True, False], [0, 1]))
>>> cdims = (['A', 'B'], [(True, 0), (True, 1),
... (False, 0), (False, 1)])
Attributes
----------
learners : list
The sequence of `~adaptive.BaseLearner`\s.
function : callable
A function that calls the functions of the underlying learners.
Its signature is ``function(learner_index, point)``.
strategy : 'loss_improvements' (default), 'loss', 'npoints', or 'cycle'.
The points that the `BalancingLearner` choses can be either based on:
the best 'loss_improvements', the smallest total 'loss' of the
child learners, the number of points per learner, using 'npoints',
or by cycling through the learners one by one using 'cycle'.
One can dynamically change the strategy while the simulation is
running by changing the ``learner.strategy`` attribute.
Notes
-----
This learner compares the `loss` calculated from the "child" learners.
This requires that the 'loss' from different learners *can be meaningfully
compared*. For the moment we enforce this restriction by requiring that
all learners are the same type but (depending on the internals of the
learner) it may be that the loss cannot be compared *even between learners
of the same type*. In this case the `~adaptive.BalancingLearner` will
behave in an undefined way. Change the `strategy` in that case.
"""
def __init__(self, learners, *, cdims=None, strategy="loss_improvements"):
self.learners = learners
# Naively we would make 'function' a method, but this causes problems
# when using executors from 'concurrent.futures' because we have to
# pickle the whole learner.
self.function = partial(dispatch, [l.function for l in self.learners])
self._ask_cache = {}
self._loss = {}
self._pending_loss = {}
self._cdims_default = cdims
if len({learner.__class__ for learner in self.learners}) > 1:
raise TypeError(
"A BalacingLearner can handle only one type" " of learners."
)
self.strategy = strategy
@property
def data(self):
data = {}
for i, l in enumerate(self.learners):
data.update({(i, p): v for p, v in l.data.items()})
return data
@property
def pending_points(self):
pending_points = set()
for i, l in enumerate(self.learners):
pending_points.update({(i, p) for p in l.pending_points})
return pending_points
@property
def npoints(self):
return sum(l.npoints for l in self.learners)
@property
def strategy(self):
"""Can be either 'loss_improvements' (default), 'loss', 'npoints', or
'cycle'. The points that the `BalancingLearner` choses can be either
based on: the best 'loss_improvements', the smallest total 'loss' of
the child learners, the number of points per learner, using 'npoints',
or by going through all learners one by one using 'cycle'.
One can dynamically change the strategy while the simulation is
running by changing the ``learner.strategy`` attribute."""
return self._strategy
@strategy.setter
def strategy(self, strategy):
self._strategy = strategy
if strategy == "loss_improvements":
self._ask_and_tell = self._ask_and_tell_based_on_loss_improvements
elif strategy == "loss":
self._ask_and_tell = self._ask_and_tell_based_on_loss
elif strategy == "npoints":
self._ask_and_tell = self._ask_and_tell_based_on_npoints
elif strategy == "cycle":
self._ask_and_tell = self._ask_and_tell_based_on_cycle
self._cycle = itertools.cycle(range(len(self.learners)))
else:
raise ValueError(
'Only strategy="loss_improvements", strategy="loss",'
' strategy="npoints", or strategy="cycle" is implemented.'
)
def _ask_and_tell_based_on_loss_improvements(self, n):
selected = [] # tuples ((learner_index, point), loss_improvement)
total_points = [l.npoints + len(l.pending_points) for l in self.learners]
for _ in range(n):
to_select = []
for index, learner in enumerate(self.learners):
# Take the points from the cache
if index not in self._ask_cache:
self._ask_cache[index] = learner.ask(n=1, tell_pending=False)
points, loss_improvements = self._ask_cache[index]
to_select.append(
((index, points[0]), (loss_improvements[0], -total_points[index]))
)
# Choose the optimal improvement.
(index, point), (loss_improvement, _) = max(to_select, key=itemgetter(1))
total_points[index] += 1
selected.append(((index, point), loss_improvement))
self.tell_pending((index, point))
points, loss_improvements = map(list, zip(*selected))
return points, loss_improvements
def _ask_and_tell_based_on_loss(self, n):
selected = [] # tuples ((learner_index, point), loss_improvement)
total_points = [l.npoints + len(l.pending_points) for l in self.learners]
for _ in range(n):
losses = self._losses(real=False)
index, _ = max(
enumerate(zip(losses, (-n for n in total_points))), key=itemgetter(1)
)
total_points[index] += 1
# Take the points from the cache
if index not in self._ask_cache:
self._ask_cache[index] = self.learners[index].ask(n=1)
points, loss_improvements = self._ask_cache[index]
selected.append(((index, points[0]), loss_improvements[0]))
self.tell_pending((index, points[0]))
points, loss_improvements = map(list, zip(*selected))
return points, loss_improvements
def _ask_and_tell_based_on_npoints(self, n):
selected = [] # tuples ((learner_index, point), loss_improvement)
total_points = [l.npoints + len(l.pending_points) for l in self.learners]
for _ in range(n):
index = np.argmin(total_points)
# Take the points from the cache
if index not in self._ask_cache:
self._ask_cache[index] = self.learners[index].ask(n=1)
points, loss_improvements = self._ask_cache[index]
total_points[index] += 1
selected.append(((index, points[0]), loss_improvements[0]))
self.tell_pending((index, points[0]))
points, loss_improvements = map(list, zip(*selected))
return points, loss_improvements
def _ask_and_tell_based_on_cycle(self, n):
points, loss_improvements = [], []
for _ in range(n):
index = next(self._cycle)
point, loss_improvement = self.learners[index].ask(n=1)
points.append((index, point[0]))
loss_improvements.append(loss_improvement[0])
self.tell_pending((index, point[0]))
return points, loss_improvements
def ask(self, n, tell_pending=True):
"""Chose points for learners."""
if n == 0:
return [], []
if not tell_pending:
with restore(*self.learners):
return self._ask_and_tell(n)
else:
return self._ask_and_tell(n)
def tell(self, x, y):
index, x = x
self._ask_cache.pop(index, None)
self._loss.pop(index, None)
self._pending_loss.pop(index, None)
self.learners[index].tell(x, y)
def tell_pending(self, x):
index, x = x
self._ask_cache.pop(index, None)
self._loss.pop(index, None)
self.learners[index].tell_pending(x)
def _losses(self, real=True):
losses = []
loss_dict = self._loss if real else self._pending_loss
for index, learner in enumerate(self.learners):
if index not in loss_dict:
loss_dict[index] = learner.loss(real)
losses.append(loss_dict[index])
return losses
@cache_latest
def loss(self, real=True):
losses = self._losses(real)
return max(losses)
def plot(self, cdims=None, plotter=None, dynamic=True):
"""Returns a DynamicMap with sliders.
Parameters
----------
cdims : sequence of dicts, or (keys, iterable of values), optional
Constant dimensions; the parameters that label the learners.
Example inputs that all give identical results:
- sequence of dicts:
>>> cdims = [{'A': True, 'B': 0},
... {'A': True, 'B': 1},
... {'A': False, 'B': 0},
... {'A': False, 'B': 1}]`
- tuple with (keys, iterable of values):
>>> cdims = (['A', 'B'], itertools.product([True, False], [0, 1]))
>>> cdims = (['A', 'B'], [(True, 0), (True, 1),
... (False, 0), (False, 1)])
plotter : callable, optional
A function that takes the learner as a argument and returns a
holoviews object. By default ``learner.plot()`` will be called.
dynamic : bool, default True
Return a `holoviews.core.DynamicMap` if True, else a
`holoviews.core.HoloMap`. The `~holoviews.core.DynamicMap` is
rendered as the sliders change and can therefore not be exported
to html. The `~holoviews.core.HoloMap` does not have this problem.
Returns
-------
dm : `holoviews.core.DynamicMap` (default) or `holoviews.core.HoloMap`
A `DynamicMap` ``(dynamic=True)`` or `HoloMap`
``(dynamic=False)`` with sliders that are defined by `cdims`.
"""
hv = ensure_holoviews()
cdims = cdims or self._cdims_default
if cdims is None:
cdims = [{"i": i} for i in range(len(self.learners))]
elif not isinstance(cdims[0], dict):
# Normalize the format
keys, values_list = cdims
cdims = [dict(zip(keys, values)) for values in values_list]
mapping = {tuple(_cdims.values()): l for l, _cdims in zip(self.learners, cdims)}
d = defaultdict(list)
for _cdims in cdims:
for k, v in _cdims.items():
d[k].append(v)
def plot_function(*args):
with suppress(KeyError):
learner = mapping[tuple(args)]
return learner.plot() if plotter is None else plotter(learner)
dm = hv.DynamicMap(plot_function, kdims=list(d.keys()))
dm = dm.redim.values(**d)
dm.cache_size = 1
if dynamic:
# XXX: change when https://github.com/pyviz/holoviews/issues/3637
# is fixed.
return dm.map(lambda obj: obj.opts(framewise=True), hv.Element)
else:
# XXX: change when https://github.com/ioam/holoviews/issues/3085
# is fixed.
vals = {d.name: d.values for d in dm.dimensions() if d.values}
return hv.HoloMap(dm.select(**vals))
def remove_unfinished(self):
"""Remove uncomputed data from the learners."""
for learner in self.learners:
learner.remove_unfinished()
@classmethod
def from_product(cls, f, learner_type, learner_kwargs, combos):
"""Create a `BalancingLearner` with learners of all combinations of
named variables’ values. The `cdims` will be set correctly, so calling
`learner.plot` will be a `holoviews.core.HoloMap` with the correct labels.
Parameters
----------
f : callable
Function to learn, must take arguments provided in in `combos`.
learner_type : `BaseLearner`
The learner that should wrap the function. For example `Learner1D`.
learner_kwargs : dict
Keyword argument for the `learner_type`. For example `dict(bounds=[0, 1])`.
combos : dict (mapping individual fn arguments -> sequence of values)
For all combinations of each argument a learner will be instantiated.
Returns
-------
learner : `BalancingLearner`
A `BalancingLearner` with learners of all combinations of `combos`
Example
-------
>>> def f(x, n, alpha, beta):
... return scipy.special.eval_jacobi(n, alpha, beta, x)
>>> combos = {
... 'n': [1, 2, 4, 8, 16],
... 'alpha': np.linspace(0, 2, 3),
... 'beta': np.linspace(0, 1, 5),
... }
>>> learner = BalancingLearner.from_product(
... f, Learner1D, dict(bounds=(0, 1)), combos)
Notes
-----
The order of the child learners inside `learner.learners` is the same
as ``adaptive.utils.named_product(**combos)``.
"""
learners = []
arguments = named_product(**combos)
for combo in arguments:
learner = learner_type(function=partial(f, **combo), **learner_kwargs)
learners.append(learner)
return cls(learners, cdims=arguments)
def save(self, fname, compress=True):
"""Save the data of the child learners into pickle files
in a directory.
Parameters
----------
fname: callable or sequence of strings
Given a learner, returns a filename into which to save the data.
Or a list (or iterable) with filenames.
compress : bool, default True
Compress the data upon saving using `gzip`. When saving
using compression, one must load it with compression too.
Example
-------
>>> def combo_fname(learner):
... val = learner.function.keywords # because functools.partial
... fname = '__'.join([f'{k}_{v}.pickle' for k, v in val.items()])
... return 'data_folder/' + fname
>>>
>>> def f(x, a, b): return a * x**2 + b
>>>
>>> learners = [Learner1D(functools.partial(f, **combo), (-1, 1))
... for combo in adaptive.utils.named_product(a=[1, 2], b=[1])]
>>>
>>> learner = BalancingLearner(learners)
>>> # Run the learner
>>> runner = adaptive.Runner(learner)
>>> # Then save
>>> learner.save(combo_fname) # use 'load' in the same way
"""
if isinstance(fname, Iterable):
for l, _fname in zip(self.learners, fname):
l.save(_fname, compress=compress)
else:
for l in self.learners:
l.save(fname(l), compress=compress)
def load(self, fname, compress=True):
"""Load the data of the child learners from pickle files
in a directory.
Parameters
----------
fname: callable or sequence of strings
Given a learner, returns a filename from which to load the data.
Or a list (or iterable) with filenames.
compress : bool, default True
If the data is compressed when saved, one must load it
with compression too.
Example
-------
See the example in the `BalancingLearner.save` doc-string.
"""
if isinstance(fname, Iterable):
for l, _fname in zip(self.learners, fname):
l.load(_fname, compress=compress)
else:
for l in self.learners:
l.load(fname(l), compress=compress)
def _get_data(self):
return [l._get_data() for l in self.learners]
def _set_data(self, data):
for l, _data in zip(self.learners, data):
l._set_data(_data)
def __getstate__(self):
return (
self.learners,
self._cdims_default,
self.strategy,
)
def __setstate__(self, state):
learners, cdims, strategy = state
self.__init__(learners, cdims=cdims, strategy=strategy)