def _update_samples_with_object_indices(
database_class: DataBase, object_indices: list,
is_queryable: bool, epoch: int) -> DataBase:
"""
Runs database class update_samples methods with object indices
Parameters
----------
database_class
An instance of DataBase class
object_indices
List of indexes identifying objects to be moved.
is_queryable
If True, consider queryable flag. Default is False.
epoch
Day since beginning of survey. Default is 20.
"""
database_class.update_samples(
object_indices, queryable=is_queryable, epoch=epoch)
return database_class
def learn_loop(nloops: int,
strategy: str,
path_to_features: str,
output_metrics_file: str,
output_queried_file: str,
features_method: str = 'Bazin',
classifier: str = 'RandomForest',
training: str = 'original',
batch: int = 1,
survey: str = 'DES',
nclass: int = 2,
photo_class_thr: float = 0.5,
photo_ids_to_file: bool = False,
photo_ids_froot: str = ' ',
classifier_bootstrap: bool = False,
save_predictions: bool = False,
sep_files=False,
pred_dir: str = None,
queryable: bool = False,
metric_label: str = 'snpcc',
save_alt_class: bool = False,
SNANA_types: bool = False,
metadata_fname: str = None,
bar: bool = True,
**kwargs):
"""
Perform the active learning loop. All results are saved to file.
Parameters
----------
nloops: int
Number of active learning loops to run.
strategy: str
Query strategy. Options are 'UncSampling', 'RandomSampling',
'UncSamplingEntropy', 'UncSamplingLeastConfident', 'UncSamplingMargin',
'QBDMI' and 'QBDEntropy'.
path_to_features: str or dict
Complete path to input features file.
if dict, keywords should be 'train' and 'test',
and values must contain the path for separate train
and test sample files.
output_metrics_file: str
Full path to output file to store metric values of each loop.
output_queried_file: str
Full path to output file to store the queried sample.
features_method: str (optional)
Feature extraction method. Currently only 'Bazin' is implemented.
classifier: str (optional)
Machine Learning algorithm.
Currently implemented options are 'RandomForest', 'GradientBoostedTrees',
'K-NNclassifier','MLPclassifier','SVMclassifier' and 'NBclassifier'.
Default is 'RandomForest'.
sep_files: bool (optional)
If True, consider train and test samples separately read
from independent files. Default is False.
batch: int (optional)
Size of batch to be queried in each loop. Default is 1.
classifier_bootstrap: bool (optional)
Flag for bootstrapping on the classifier
Must be true if using disagreement based strategy.
metadata_fname: str (optional)
Complete path to PLAsTiCC zenodo test metadata. Only used it
SNANA_types == True. Default is None.
metric_label: str (optional)
Choice of metric.
Currently only "snpcc", "cosmo" or "snpcc_cosmo" are accepted.
Default is "snpcc".
nclass: int (optional)
Number of classes to consider in the classification
Currently only nclass == 2 is implemented.
photo_class_thr: float (optional)
Threshold for photometric classification. Default is 0.5.
Only used if photo_ids is True.
photo_ids_to_file: bool (optional)
If True, save photometric ids to file. Default is False.
photo_ids_froot: str (optional)
Output root of file name to store photo ids.
Only used if photo_ids is True.
pred_dir: str (optional)
Output diretory to store prediction file for each loop.
Only used if `save_predictions==True`.
queryable: bool (optional)
If True, check if randomly chosen object is queryable.
Default is False.
save_alt_class: bool (optional)
If True, train the model and save classifications for alternative
query label (this is necessary to calculate impact on cosmology).
Default is False.
save_predictions: bool (optional)
If True, save classification predictions to file in each loop.
Default is False.
SNANA_types: bool (optional)
If True, translate zenodo types to SNANA codes.
Default is False.
survey: str (optional)
'DES' or 'LSST'. Default is 'DES'.
Name of the survey which characterizes filter set.
training: str or int (optional)
Choice of initial training sample.
If 'original': begin from the train sample flagged in the file
If int: choose the required number of samples at random,
ensuring that at least half are SN Ia
Default is 'original'.
bar: bool (optional)
If True, display progress bar.
kwargs: extra parameters
All keywords required by the classifier function.
"""
if 'QBD' in strategy and not classifier_bootstrap:
raise ValueError(
'Bootstrap must be true when using disagreement strategy')
# initiate object
database_class = DataBase()
logging.info('Loading features')
database_class = load_features(database_class, path_to_features, survey,
features_method, nclass, training,
queryable, sep_files)
logging.info('Running active learning loop')
if bar:
ensemble = progressbar.progressbar(range(nloops))
else:
ensemble = range(nloops)
for iteration_step in ensemble:
if not bar:
print(iteration_step)
database_class = run_classification(database_class, classifier,
classifier_bootstrap, pred_dir,
save_predictions, iteration_step,
**kwargs)
run_evaluation(database_class, metric_label)
save_photo_ids(database_class, photo_ids_to_file, SNANA_types,
metadata_fname, photo_class_thr, iteration_step,
photo_ids_froot, '.dat')
indices_to_query = run_make_query(database_class, strategy, batch,
queryable)
if save_alt_class and batch == 1:
database_class_alternative = copy.deepcopy(database_class)
database_class_alternative = update_alternative_label(
database_class_alternative, indices_to_query, iteration_step,
classifier, pred_dir, save_predictions, metric_label,
SNANA_types, photo_ids_to_file, metadata_fname,
photo_class_thr, photo_ids_froot, **kwargs)
_save_metrics_and_queried_samples(database_class_alternative,
output_metrics_file,
output_queried_file,
iteration_step, batch, False,
'_alt_label.dat')
elif save_alt_class and batch > 1:
raise ValueError('Alternative label only works with batch=1!')
database_class.update_samples(indices_to_query,
epoch=iteration_step,
queryable=queryable,
alternative_label=False)
_save_metrics_and_queried_samples(database_class, output_metrics_file,
output_queried_file, iteration_step,
batch, False)
return database_class
def update_alternative_label(
database_class_alternative: DataBase, indices_to_query: list,
iteration_step: int, classifier: str, pred_dir: str,
is_save_prediction: bool, metric_label: str, is_save_snana_types: bool,
is_save_photoids_to_file: bool, meta_data_fname: str,
photo_class_threshold: float, photo_ids_froot: str, **kwargs: dict):
"""
Function to update active learning training with alternative label
Parameters
----------
database_class_alternative
An instance of DataBase class for alternative label
indices_to_query
List of indexes identifying objects to be moved.
iteration_step
active learning iteration number
classifier
Machine Learning algorithm.
Currently implemented options are 'RandomForest', 'GradientBoostedTrees',
'K-NNclassifier','MLPclassifier','SVMclassifier' and 'NBclassifier'.
Default is 'RandomForest'.
pred_dir
Output diretory to store prediction file for each loop.
Only used if `save_predictions==True`.
is_save_prediction
if predictions should be saved
metric_label
Choice of metric.
Currently only "snpcc", "cosmo" or "snpcc_cosmo" are accepted.
Default is "snpcc".
is_save_snana_types
if True, translate type to SNANA codes and
add column with original values. Default is False.
is_save_photoids_to_file
If true, populate the photo_Ia_list attribute. Otherwise
write to file. Default is False.
meta_data_fname
Full path to PLAsTiCC zenodo test metadata file.
photo_class_threshold
Probability threshold above which an object is considered Ia.
photo_ids_froot
Output root of file name to store photo ids.
Only used if photo_ids is True.
kwargs
additional arguments
"""
database_class_alternative.update_samples(indices_to_query,
epoch=iteration_step,
alternative_label=True)
database_class_alternative = run_classification(database_class_alternative,
classifier, False,
pred_dir,
is_save_prediction,
iteration_step, **kwargs)
run_evaluation(database_class_alternative, metric_label)
save_photo_ids(database_class_alternative, is_save_photoids_to_file,
is_save_snana_types, meta_data_fname, photo_class_threshold,
iteration_step, photo_ids_froot, '_alt_label.dat')
return database_class_alternative
def learn_loop(nloops: int, strategy: str, path_to_features: str,
output_diag_file: str, output_queried_file: str,
features_method='Bazin', classifier='RandomForest',
training='original', batch=1, screen=True, survey='DES',
perc=0.1, nclass=2):
"""Perform the active learning loop. All results are saved to file.
Parameters
----------
nloops: int
Number of active learning loops to run.
strategy: str
Query strategy. Options are 'UncSampling' and 'RandomSampling'.
path_to_features: str or dict
Complete path to input features file.
if dict, keywords should be 'train' and 'test',
and values must contain the path for separate train
and test sample files.
output_diag_file: str
Full path to output file to store diagnostics of each loop.
output_queried_file: str
Full path to output file to store the queried sample.
features_method: str (optional)
Feature extraction method. Currently only 'Bazin' is implemented.
classifier: str (optional)
Machine Learning algorithm.
Currently only 'RandomForest' is implemented.
training: str or int (optional)
Choice of initial training sample.
If 'original': begin from the train sample flagged in the file
If int: choose the required number of samples at random,
ensuring that at least half are SN Ia
Default is 'original'.
batch: int (optional)
Size of batch to be queried in each loop. Default is 1.
screen: bool (optional)
If True, print on screen number of light curves processed.
survey: str (optional)
'DES' or 'LSST'. Default is 'DES'.
Name of the survey which characterizes filter set.
perc: float in [0,1] (optioal)
Percentile chosen to identify the new query.
Only used for PercentileSampling.
Default is 0.1.
nclass: int (optional)
Number of classes to consider in the classification
Currently only nclass == 2 is implemented.
"""
## This module will need to be expanded for RESSPECT
# initiate object
data = DataBase()
# load features
if isinstance(path_to_features, str):
data.load_features(path_to_features, method=features_method,
screen=screen, survey=survey)
# separate training and test samples
data.build_samples(initial_training=training, nclass=nclass)
else:
data.load_features(path_to_features['train'], method=features_method,
screen=screen, survey=survey, sample='train')
data.load_features(path_to_features['test'], method=features_method,
screen=screen, survey=survey, sample='test')
data.build_samples(initial_training=training, nclass=nclass,
screen=screen, sep_files=True)
for loop in range(nloops):
if screen:
print('Processing... ', loop)
# classify
data.classify(method=classifier)
# calculate metrics
data.evaluate_classification()
# choose object to query
indx = data.make_query(strategy=strategy, batch=batch, perc=perc)
# update training and test samples
data.update_samples(indx, loop=loop)
# save diagnostics for current state
data.save_metrics(loop=loop, output_metrics_file=output_diag_file,
batch=batch, epoch=loop)
# save query sample to file
data.save_queried_sample(output_queried_file, loop=loop,
full_sample=False)
def time_domain_loop(days: list, output_diag_file: str,
output_queried_file: str,
path_to_features_dir: str, strategy: str,
batch=1, canonical = False, classifier='RandomForest',
features_method='Bazin', path_to_canonical="",
path_to_full_lc_features="",
screen=True, training='original'):
"""Perform the active learning loop. All results are saved to file.
Parameters
----------
days: list
List of 2 elements. First and last day of observations since the
beginning of the survey.
output_diag_file: str
Full path to output file to store diagnostics of each loop.
output_queried_file: str
Full path to output file to store the queried sample.
path_to_features_dir: str
Complete path to directory holding features files for all days.
strategy: str
Query strategy. Options are 'UncSampling' and 'RandomSampling'.
batch: int (optional)
Size of batch to be queried in each loop. Default is 1.
canonical: bool (optional)
If True, restrict the search to the canonical sample.
classifier: str (optional)
Machine Learning algorithm.
Currently only 'RandomForest' is implemented.
features_method: str (optional)
Feature extraction method. Currently only 'Bazin' is implemented.
path_to_canonical: str (optional)
Path to canonical sample features files.
It is only used if "strategy==canonical".
path_to_full_lc_features: str (optional)
Path to full light curve features file.
Only used if training is a number.
screen: bool (optional)
If True, print on screen number of light curves processed.
training: str or int (optional)
Choice of initial training sample.
If 'original': begin from the train sample flagged in the file
If int: choose the required number of samples at random,
ensuring that at least half are SN Ia
Default is 'original'.
"""
## This will need to change for RESSPECT
# initiate object
data = DataBase()
# load features for the first day
path_to_features = path_to_features_dir + 'day_' + str(int(days[0])) + '.dat'
data.load_features(path_to_features, method=features_method,
screen=screen)
# change training
if training == 'original':
data.build_samples(initial_training='original')
full_lc_features = get_original_training(path_to_features=path_to_full_lc_features)
data.train_metadata = full_lc_features.train_metadata
data.train_labels = full_lc_features.train_labels
data.train_features = full_lc_features.train_features
else:
data.build_samples(initial_training=int(training))
# get list of canonical ids
if canonical:
canonical = DataBase()
canonical.load_features(path_to_file=path_to_canonical)
data.queryable_ids = canonical.queryable_ids
for night in range(int(days[0]), int(days[-1]) - 1):
if screen:
print('Processing night: ', night)
# cont loop
loop = night - int(days[0])
# classify
data.classify(method=classifier)
# calculate metrics
data.evaluate_classification()
# choose object to query
indx = data.make_query(strategy=strategy, batch=batch)
# update training and test samples
data.update_samples(indx, loop=loop)
# save diagnostics for current state
data.save_metrics(loop=loop, output_metrics_file=output_diag_file,
batch=batch, epoch=night)
# save query sample to file
data.save_queried_sample(output_queried_file, loop=loop,
full_sample=False)
# load features for next day
path_to_features2 = path_to_features_dir + 'day_' + str(night + 1) + '.dat'
data_tomorrow = DataBase()
data_tomorrow.load_features(path_to_features2, method=features_method,
screen=False)
# identify objects in the new day which must be in training
train_flag = np.array([item in data.train_metadata['id'].values
for item in data_tomorrow.metadata['id'].values])
# use new data
data.train_metadata = data_tomorrow.metadata[train_flag]
data.train_features = data_tomorrow.features.values[train_flag]
data.test_metadata = data_tomorrow.metadata[~train_flag]
data.test_features = data_tomorrow.features.values[~train_flag]
# new labels
data.train_labels = np.array([int(item == 'Ia') for item in
data.train_metadata['type'].values])
data.test_labels = np.array([int(item == 'Ia') for item in
data.test_metadata['type'].values])
if strategy == 'canonical':
data.queryable_ids = canonical.queryable_ids
if queryable:
queryable_flag = data_tomorrow.metadata['queryable'].values
queryable_test_flag = np.logical_and(~train_flag, queryable_flag)
data.queryable_ids = data_tomorrow.metadata['id'].values[queryable_test_flag]
else:
data.queryable_ids = data_tomorrow.metadata['id'].values[~train_flag]
if screen:
print('Training set size: ', data.train_metadata.shape[0])
print('Test set size: ', data.test_metadata.shape[0])