def gridSearch(options, use_datasets, numExamples, compute_mistakes=False, verbose=False, parallelize=False):
if MODEL_KEYWORD not in options:
print 'ERROR: must specify models for grid search under "%s" key.' % (MODEL_KEYWORD)
return
paramCombos = myProduct(options)
partialTestCombo = partial(testCombo, use_datasets=use_datasets, numExamples=numExamples, compute_mistakes=compute_mistakes, verbose=verbose)
if parallelize:
from pathos.multiprocessing import Pool
p = Pool(5)
try:
result = p.map_async(partialTestCombo, paramCombos)
result = result.get(999999999)
bestScore, bestParamsStr, bestCombo = max(result, key=lambda x:x[0])
sys.stdout = open("best.out", "w")
print 'Best score of %s was achieved by parameters:\n%s' % (bestScore, bestParamsStr)
except KeyboardInterrupt:
p.terminate()
print "You cancelled the program!"
sys.exit(1)
else:
bestScore, bestCombo, bestComboStr = float('-inf'), None, ''
for paramCombo in paramCombos:
score, paramsStr, _ = testCombo(paramCombo, use_datasets=use_datasets, numExamples=numExamples, compute_mistakes=compute_mistakes, verbose=verbose, parallelize=False)
if score > bestScore:
bestScore, bestCombo, bestComboStr = score, paramCombo, paramsStr
print 'Best score of %s was achieved by parameters:\n%s' % (bestScore, bestComboStr)
def create_initial_population(self):
"""Create members of the first population randomly."""
for _ in range(self.pop_size):
individual = Particle(self.chromosome_size, self.fitness_function)
if not self.pool:
individual.calculate_fitness()
self.add_individual_to_pop(individual)
if self.pool:
p = Pool(self.pool_size)
manager = Manager()
lock = manager.Lock()
counter = manager.Value('i', 0)
def pool_function(inside_lock, inside_counter, inside_member):
inside_lock.acquire()
inside_counter.value += 1
inside_lock.release()
fitness_value = inside_member.calculate_fitness(
gpu=inside_counter.value % 4)
return fitness_value
func = partial(pool_function, lock, counter)
fitness_values = p.map(func, self.current_population[:])
for value, member in zip(fitness_values,
self.current_population[:]):
member.fitness = value
p.terminate()
def run_mcmc(self):
complete = self.chains.check_completness()
if not complete:
for i in range(self.nchains):
self.chains.chains.append(
Chain(self.chains.chains_filename, self.covfile, nchain=i, nsteps=self.nsteps))
pool = Pool(processes=self.nchains)
try:
# Without the .get(9999), you can't interrupt this with Ctrl+C.
pool.map_async(self.mcmc, self.chains.chains).get(999999)
pool.close()
pool.join()
# to skip lines after the progress bars
print '\n' * self.nchains
except KeyboardInterrupt:
pool.terminate()
self.likelihood = self.chains.chains_to_likelihood()
self.likelihood.stats(self.covfile)
# [self.results[i].append(self.likelihood.pdfs[i].mean) for i in range(self.chains.dim)]
# self.p = [self.likelihood.pdfs[i].mean for i in range(self.chains.dim)]
self.p = self.chains.best_row_params
self.simulation(self.spectrum.lambdas, *self.p)
# [self.results_err[i].append([self.likelihood.pdfs[i].error_high,self.likelihood.pdfs[i].error_low]) for i in range(self.chains.dim)]
# if(self.plot):
self.likelihood.triangle_plots()
self.plot_fit()
# if convergence_test :
self.chains.convergence_tests()
return self.likelihood
def vw_train_and_test(self, options_list, data_file_paths):
def init_worker():
signal.signal(signal.SIGINT, signal.SIG_IGN)
def run_learner(options):
train_options = {
'final_regressor': os.path.join(self.work_dir, id_generator()),
'data': data_file_paths['train'],
'cache_file': data_file_paths['train'] + '.cache'
}
train_options.update(options)
test_options = {
'data': data_file_paths['test'],
'predictions': os.path.join(self.work_dir, id_generator()),
'cache_file': data_file_paths['test'] + '.cache'
}
test_options.update(options)
# TO DO: remove below if.
if 'kill_cache' in options:
del train_options['kill_cache']
del test_options['kill_cache']
vw_wrapper = VW_Wrapper(verbose=False)
vw_wrapper.train(train_options)
predictions = vw_wrapper.test(test_options)
os.remove(train_options['final_regressor'])
os.remove(test_options['predictions'])
return options, predictions
if len(options_list) > 1:
try:
if not os.path.isfile(data_file_paths['test'] +
'.cache') or not os.path.isfile(
data_file_paths['train'] + '.cache'):
run_learner(options_list[0])
pool = Pool(len(options_list), init_worker)
result_list = pool.map_async(run_learner,
options_list).get(99999999)
pool.close()
pool.join()
return result_list
except KeyboardInterrupt:
print ' Keyboard Interrupt, exiting...) '
pool.terminate()
pool.join()
sys.exit(0)
elif len(options_list) == 1:
return [run_learner(options_list[0])]
else:
return []
def gridSearch(options,
use_datasets,
numExamples,
compute_mistakes=False,
verbose=False,
parallelize=False):
if MODEL_KEYWORD not in options:
print 'ERROR: must specify models for grid search under "%s" key.' % (
MODEL_KEYWORD)
return
paramCombos = myProduct(options)
partialTestCombo = partial(testCombo,
use_datasets=use_datasets,
numExamples=numExamples,
compute_mistakes=compute_mistakes,
verbose=verbose)
if parallelize:
from pathos.multiprocessing import Pool
p = Pool(5)
try:
result = p.map_async(partialTestCombo, paramCombos)
result = result.get(999999999)
bestScore, bestParamsStr, bestCombo = max(result,
key=lambda x: x[0])
sys.stdout = open("best.out", "w")
print 'Best score of %s was achieved by parameters:\n%s' % (
bestScore, bestParamsStr)
except KeyboardInterrupt:
p.terminate()
print "You cancelled the program!"
sys.exit(1)
else:
bestScore, bestCombo, bestComboStr = float('-inf'), None, ''
for paramCombo in paramCombos:
score, paramsStr, _ = testCombo(paramCombo,
use_datasets=use_datasets,
numExamples=numExamples,
compute_mistakes=compute_mistakes,
verbose=verbose,
parallelize=False)
if score > bestScore:
bestScore, bestCombo, bestComboStr = score, paramCombo, paramsStr
print 'Best score of %s was achieved by parameters:\n%s' % (
bestScore, bestComboStr)
def partial_dependence(self, feature_ids, modelinstance, filter_classes=None, grid=None,
grid_resolution=30, n_jobs=-1, grid_range=None, sample=True,
sampling_strategy='random-choice', n_samples=1000,
bin_count=50, samples_per_bin=10, return_metadata=False):
"""
Approximates the partial dependence of the predict_fn with respect to the
variables passed.
Parameters:
-----------
feature_ids: list
the names/ids of the features for which partial dependence is to be computed.
Note that the algorithm's complexity scales exponentially with additional
features, so generally one should only look at one or two features at a
time. These feature ids must be available in the class's associated DataSet.
As of now, we only support looking at 1 or 2 features at a time.
modelinstance: skater.model.model.Model subtype
an estimator function of a fitted model used to derive prediction. Supports
classification and regression. Supports classification(binary, multi-class) and regression.
predictions = predict_fn(data)
Can either by a skater.model.remote.DeployedModel or a
skater.model.local.InMemoryModel
filter_classes: array type
The classes to run partial dependence on. Default None invokes all classes.
Only used in classification models.
grid: numpy.ndarray
2 dimensional array on which we fix values of features. Note this is
determined automatically if not given based on the percentiles of the
dataset.
grid_resolution: int
how many unique values to include in the grid. If the percentile range
is 5% to 95%, then that range will be cut into <grid_resolution>
equally size bins. Defaults to 30.
n_jobs: int
The number of CPUs to use to compute the PDs. -1 means 'all CPUs'.
Defaults to using all cores(-1).
grid_range: tuple
the percentile extrama to consider. 2 element tuple, increasing, bounded
between 0 and 1.
sample: boolean
Whether to sample from the original dataset.
sampling_strategy: string
If sampling, which approach to take. See DataSet.generate_sample for
details.
n_samples: int
The number of samples to use from the original dataset. Note this is
only active if sample = True and sampling strategy = 'uniform'. If
using 'uniform-over-similarity-ranks', use samples per bin
bin_count: int
The number of bins to use when using the similarity based sampler. Note
this is only active if sample = True and
sampling_strategy = 'uniform-over-similarity-ranks'.
total samples = bin_count * samples per bin.
samples_per_bin: int
The number of samples to collect for each bin within the sampler. Note
this is only active if sample = True and
sampling_strategy = 'uniform-over-similarity-ranks'. If using
sampling_strategy = 'uniform', use n_samples.
total samples = bin_count * samples per bin.
return_metadata: boolean
:Example:
>>> from skater.model import InMemoryModel
>>> from skater.core.explanations import Interpretation
>>> from sklearn.ensemble import RandomForestClassier
>>> from sklearn.datasets import load_boston
>>> boston = load_boston()
>>> X = boston.data
>>> y = boston.target
>>> features = boston.feature_names
>>> rf = RandomForestClassier()
>>> rf.fit(X,y)
>>> model = InMemoryModel(rf, examples = X)
>>> interpreter = Interpretation()
>>> interpreter.load_data(X)
>>> feature_ids = ['ZN','CRIM']
>>> interpreter.partial_dependence.partial_dependence(features,model)
"""
if self.data_set is None:
load_data_not_called_err_msg = "self.interpreter.data_set not found. " \
"Please call Interpretation.load_data " \
"before running this method."
raise(exceptions.DataSetNotLoadedError(load_data_not_called_err_msg))
feature_ids = self._check_features(feature_ids)
if filter_classes:
err_msg = "members of filter classes must be" \
"members of modelinstance.classes." \
"Expected members of: " \
"{0}\n" \
"got: " \
"{1}".format(modelinstance.target_names,
filter_classes)
assert all([i in modelinstance.target_names for i in filter_classes]), err_msg
# TODO: There might be a better place to do this check
if not isinstance(modelinstance, ModelType):
raise(exceptions.ModelError("Incorrect estimator function used for computing partial dependence, try one "
"creating one with skater.model.local.InMemoryModel or"
"skater.model.remote.DeployedModel"))
if modelinstance.model_type == 'classifier' and modelinstance.probability is False:
if modelinstance.unique_values is None:
raise(exceptions.ModelError('If using classifier without probability scores, unique_values cannot '
'be None'))
self.interpreter.logger.warn("Classifiers with probability scores can be explained "
"more granularly than those without scores. If a prediction method with "
"scores is available, use that instead.")
# TODO: This we can change easily to functional style
missing_feature_ids = []
for feature_id in feature_ids:
if feature_id not in self.data_set.feature_ids:
missing_feature_ids.append(feature_id)
if missing_feature_ids:
missing_feature_id_err_msg = "Features {0} not found in " \
"Interpretation.data_set.feature_ids" \
"{1}".format(missing_feature_ids, self.data_set.feature_ids)
raise(KeyError(missing_feature_id_err_msg))
if grid_range is None:
grid_range = (.05, 0.95)
else:
if not hasattr(grid_range, "__iter__"):
err_msg = "Grid range {} needs to be an iterable".format(grid_range)
raise(exceptions.MalformedGridRangeError(err_msg))
self._check_grid_range(grid_range)
if not modelinstance.has_metadata:
examples = self.data_set.generate_sample(strategy='random-choice',
sample=True,
n_samples_from_dataset=10)
examples = DataManager(examples, feature_names=self.data_set.feature_ids)
modelinstance._build_model_metadata(examples)
# if you dont pass a grid, build one.
grid = np.array(grid)
if not grid.any():
# Currently, if a given feature has fewer unique values than the value
# of grid resolution, then the grid will be set to those unique values.
# Otherwise it will take the percentile
# range according with grid_resolution bins.
grid = self.data_set.generate_grid(feature_ids,
grid_resolution=grid_resolution,
grid_range=grid_range)
else:
# want to ensure all grids have 2 axes
if len(grid.shape) == 1 and \
(StaticTypes.data_types.is_string(grid[0]) or StaticTypes.data_types.is_numeric(grid[0])):
grid = grid[:, np.newaxis].T
grid_resolution = grid.shape[1]
self.interpreter.logger.debug("Grid shape used for pdp: {}".format(grid.shape))
self.interpreter.logger.debug("Grid resolution for pdp: {}".format(grid_resolution))
# make sure data_set module is giving us correct data structure
self._check_grid(grid, feature_ids)
# generate data
data_sample = self.data_set.generate_sample(strategy=sampling_strategy,
sample=sample,
n_samples_from_dataset=n_samples,
samples_per_bin=samples_per_bin,
bin_count=bin_count)
_pdp_metadata = self._build_metadata_dict(modelinstance, feature_ids, self.data_set.feature_ids, filter_classes)
self.interpreter.logger.debug("Shape of sampled data: {}".format(data_sample.shape))
self.interpreter.logger.debug("Feature Ids: {}".format(feature_ids))
self.interpreter.logger.debug("PD metadata: {}".format(_pdp_metadata))
# cartesian product of grid
grid_expanded = pd.DataFrame(list(product(*grid))).values
if grid_expanded.shape[0] <= 0:
empty_grid_expanded_err_msg = "Must have at least 1 pdp value" \
"grid shape: {}".format(grid_expanded.shape)
raise(exceptions.MalformedGridError(empty_grid_expanded_err_msg))
predict_fn = modelinstance._get_static_predictor()
n_jobs = None if n_jobs < 0 else n_jobs
pd_func = functools.partial(_compute_pd,
estimator_fn=predict_fn,
grid_expanded=grid_expanded,
pd_metadata=_pdp_metadata,
input_data=data_sample,
filter_classes=filter_classes)
arg_list = [i for i in range(grid_expanded.shape[0])]
executor_instance = Pool(n_jobs)
try:
pd_list = executor_instance.map(pd_func, arg_list)
except:
self.interpreter.logger.debug("Multiprocessing failed, going single process")
pd_list = map(pd_func, arg_list)
finally:
executor_instance.close()
executor_instance.join()
executor_instance.terminate()
if return_metadata:
return pd.DataFrame(list(pd_list)), _pdp_metadata
else:
return pd.DataFrame(list(pd_list))
def bayesian_optimisation(patience=-1,
n_iters=-1,
sample_loss=None,
bounds=None,
x0=None,
n_pre_samples=5,
gp_params=None,
random_search=False,
alpha=1e-5,
epsilon=1e-7,
pool_size=1):
""" bayesian_optimisation
Uses Gaussian Processes to optimise the loss function `sample_loss`.
Arguments:
----------
patience: integer
Number of nonimproved iterations before exit.
n_iters: integer.
Number of iterations to run the search algorithm.
sample_loss: function.
Function to be optimised.
bounds: array-like, shape = [n_params, 2].
Lower and upper bounds on the parameters of the function `sample_loss`.
x0: array-like, shape = [n_pre_samples, n_params].
Array of initial points to sample the loss function for. If None, randomly
samples from the loss function.
n_pre_samples: integer.
If x0 is None, samples `n_pre_samples` initial points from the loss function.
gp_params: dictionary.
Dictionary of parameters to pass on to the underlying Gaussian Process.
random_search: integer.
Flag that indicates whether to perform random search or L-BFGS-B optimisation
over the acquisition function.
alpha: double.
Variance of the error term of the GP.
epsilon: double.
Precision tolerance for floats.
"""
x_list = []
y_list = []
iteration = 0
no_improvement = 0
best_fitness = 100
n_params = bounds.shape[0]
print("Create {} initial points".format(n_pre_samples))
if x0 is None:
for params in np.random.uniform(bounds[:, 0], bounds[:, 1],
(n_pre_samples, bounds.shape[0])):
x_list.append(params)
if pool_size > 1:
p = Pool(pool_size)
losses = p.map(sample_loss, x_list)
y_list = losses
p.terminate()
else:
for params in x_list:
y_list.append(sample_loss(params))
else:
for params in x0:
x_list.append(params)
if pool_size > 1:
p = Pool(pool_size)
losses = p.map(sample_loss, x_list)
y_list = losses
p.terminate()
else:
for params in x_list:
y_list.append(sample_loss(params))
xp = np.array(x_list)
yp = np.array(y_list)
# Create the GP
if gp_params is not None:
model = gp.GaussianProcessRegressor(**gp_params)
else:
kernel = gp.kernels.Matern()
model = gp.GaussianProcessRegressor(kernel=kernel,
alpha=alpha,
n_restarts_optimizer=10,
normalize_y=True)
print("Use bayesian optimization to sample points.")
while True:
print("Iteration: {}".format(iteration + 1))
model.fit(xp, yp)
# Sample next hyperparameter
if random_search:
x_random = np.random.uniform(bounds[:, 0],
bounds[:, 1],
size=(random_search, n_params))
ei = -1 * expected_improvement(x_random,
model,
yp,
greater_is_better=False,
n_params=n_params)
next_sample = x_random[np.argmax(ei), :]
else:
next_sample = sample_next_hyperparameter(expected_improvement,
model,
yp,
greater_is_better=False,
bounds=bounds,
n_restarts=100)
# Duplicates will break the GP. In case of a duplicate, we will randomly sample a next query point.
if np.any(np.abs(next_sample - xp) <= epsilon):
next_sample = np.random.uniform(bounds[:, 0], bounds[:, 1],
bounds.shape[0])
# Sample loss for new set of parameters
cv_score = sample_loss(next_sample)
# Update lists
x_list.append(next_sample)
y_list.append(cv_score)
# Update xp and yp
xp = np.array(x_list)
yp = np.array(y_list)
if cv_score < best_fitness:
best_fitness = cv_score
no_improvement = 0
else:
no_improvement += 1
iteration += 1
if iteration == n_iters or patience == no_improvement:
break
return xp, yp
def add_new_individuals_function(self, name):
"""
Add new individuals to the population with a given method
(crossover, differential evolution, invasive weed, add pure individual).
"""
start = time.time()
if name == "Crossover":
current_function = partial(self.population.crossover, self.selection_function)
name = name[:9] + ' ' + self.config["selection_type"]
iterator = range(self.num_of_crossover)
elif name == "Differential evolution":
current_function = partial(self.population.differential_evolution, self.config["CR"], self.config["F"])
iterator = range(len(self.population))
elif name == "Invasive weed":
current_function = partial(self.population.invasive_weed, self.iteration, self.config["iter_max"],
self.config["e"], self.config["sigma_init"], self.config["sigma_fin"],
self.config["N_min"], self.config["N_max"])
iterator = self.population[:]
elif name == "Add pure new":
current_function = self.population.add_new_individual
iterator = range(self.num_of_new_individual)
else:
raise NameError("Bad type of function.")
if self.pool:
p = Pool(self.pool_size)
manager = Manager()
lock = manager.Lock()
counter = manager.Value('i', 0)
def pool_function(inside_lock, inside_counter, inside_member):
inside_lock.acquire()
inside_counter.value += 1
inside_lock.release()
inside_members = current_function(inside_member, gpu=inside_counter.value % 4)
return inside_members
func = partial(pool_function, lock, counter)
members = p.map(func, iterator)
if name == "Differential evolution":
self.population.current_population = members
else:
try:
members = sum(members, [])
except TypeError:
pass
for member in members:
self.population.add_individual_to_pop(member)
p.terminate()
else:
members = []
for argument in iterator:
member = current_function(argument, gpu=0)
members.append(member)
if name == "Differential evolution":
self.population.current_population = members
else:
try:
members = sum(members, [])
except TypeError:
pass
for member in members:
self.population.add_individual_to_pop(member)
step_time = time.time() - start
if step_time < 120:
print('{0} time: {1:.2f}s\n'.format(name, step_time))
else:
print('{0} time: {1:.2f}min\n'.format(name, step_time // 60))
return step_time, name
def modify_one_by_one_function(self, name):
"""Apply a function (local search, mutation) to all chromosomes."""
start = time.time()
if self.progress_bar:
print("{}:".format(name))
if name == "Local search":
current_function = self.memetic_function
elif name == "Mutation":
current_function = self.mutation_function
name = name[:8] + ' ' + self.config["mutation_type"]
else:
raise NameError("Bad type of function.")
if self.iteration > 1:
if name in self.logs[-2].keys():
if self.logs[-2][name]["step_time"] < 4:
self.progress_bar = False
else:
self.progress_bar = True
if self.fitness_function.name in ["fully connected", "convnet"]:
self.progress_bar = False
if self.pool:
p = Pool(self.pool_size)
manager = Manager()
lock = manager.Lock()
counter = manager.Value('i', 0)
if self.progress_bar:
pbar = ProgressBar(widgets=[Percentage(), Bar(), ETA()], term_width=60, maxval=len(self.population)).start()
else:
pbar = None
def pool_function(inside_lock, inside_counter, inside_member):
inside_lock.acquire()
inside_counter.value += 1
inside_lock.release()
inside_member.apply_on_chromosome(current_function, gpu=inside_counter.value % 4)
inside_lock.acquire()
if pbar:
pbar.update(inside_counter.value)
inside_lock.release()
return inside_member
func = partial(pool_function, lock, counter)
first = 1 if self.elitism and name == "Mutation" else 0
members = p.map(func, self.population[first:])
if self.elitism and name == "Mutation":
members.append(self.population[0])
self.population.current_population = members
p.terminate()
else:
if self.progress_bar:
pbar = ProgressBar(widgets=[Percentage(), Bar(), ETA()], term_width=60, maxval=len(self.population)).start()
ignor_first = self.elitism and name == "Mutation"
for i, member in enumerate(self.population):
if self.progress_bar:
pbar.update(i + 1)
if not ignor_first:
member.apply_on_chromosome(current_function)
ignor_first = False
if self.progress_bar:
pbar.finish()
step_time = time.time() - start
if step_time < 120:
print('{0} time: {1:.2f}s\n'.format(name, step_time))
else:
print('{0} time: {1:.2f}min\n'.format(name, step_time // 60))
return step_time, name
