208 lines
10 KiB
Python
208 lines
10 KiB
Python
from pathlib import Path
|
|
|
|
from sklearn.linear_model import LogisticRegression
|
|
from copy import deepcopy as cp
|
|
import quapy as qp
|
|
from _bayeisan_kdey import BayesianKDEy
|
|
from _bayesian_mapls import BayesianMAPLS
|
|
from commons import experiment_path, KDEyCLR, RESULT_DIR, MockClassifierFromPosteriors
|
|
# import datasets
|
|
from datasets import LeQuaHandler, UCIMulticlassHandler, DatasetHandler, MNISTHandler
|
|
from temperature_calibration import temp_calibration
|
|
from build.lib.quapy.data import LabelledCollection
|
|
from quapy.method.aggregative import DistributionMatchingY as DMy, AggregativeQuantifier, EMQ, CC
|
|
from quapy.model_selection import GridSearchQ
|
|
from quapy.data import Dataset
|
|
from quapy.method.confidence import BayesianCC, AggregativeBootstrap
|
|
from quapy.method.aggregative import KDEyML, ACC
|
|
from quapy.protocol import UPP
|
|
import numpy as np
|
|
from tqdm import tqdm
|
|
from collections import defaultdict
|
|
from time import time
|
|
|
|
|
|
|
|
def methods(data_handler: DatasetHandler):
|
|
"""
|
|
Returns a tuple (name, quantifier, hyperparams, bayesian/bootstrap_constructor), where:
|
|
- name: is a str representing the name of the method (e.g., 'BayesianKDEy')
|
|
- quantifier: is the base model (e.g., KDEyML())
|
|
- hyperparams: is a dictionary for the quantifier (e.g., {'bandwidth': [0.001, 0.005, 0.01, 0.05, 0.1, 0.2]})
|
|
- bayesian/bootstrap_constructor: is a function that instantiates the bayesian o bootstrap method with the
|
|
quantifier with optimized hyperparameters
|
|
"""
|
|
if isinstance(data_handler, MNISTHandler):
|
|
Cls = MockClassifierFromPosteriors
|
|
cls_hyper = {}
|
|
val_split = data_handler.get_validation().Xy # use this specific collection
|
|
else:
|
|
Cls = LogisticRegression
|
|
cls_hyper = {'classifier__C': np.logspace(-4,4,9), 'classifier__class_weight': ['balanced', None]}
|
|
val_split = 5 # k-fold cross-validation
|
|
acc_hyper = cls_hyper
|
|
# emq_hyper = {'calib': ['nbvs', 'bcts', 'ts', 'vs'], **cls_hyper}
|
|
hdy_hyper = {'nbins': [3,4,5,8,16,32], **cls_hyper}
|
|
kdey_hyper = {'bandwidth': np.logspace(-3, -1, 10), **cls_hyper}
|
|
kdey_hyper_clr = {'bandwidth': np.logspace(-2, 2, 10), **cls_hyper}
|
|
|
|
multiclass_method = 'multiclass'
|
|
only_binary = 'only_binary'
|
|
only_multiclass = 'only_multiclass'
|
|
|
|
# surrogate quantifiers
|
|
acc = ACC(Cls(), val_split=val_split)
|
|
hdy = DMy(Cls(), val_split=val_split)
|
|
kde_gau = KDEyML(Cls(), val_split=val_split)
|
|
kde_ait = KDEyCLR(Cls(), val_split=val_split)
|
|
emq = EMQ(Cls(), exact_train_prev=False, val_split=val_split)
|
|
|
|
|
|
# Bootstrap approaches:
|
|
# --------------------------------------------------------------------------------------------------------
|
|
#yield 'BootstrapACC', acc, acc_hyper, lambda hyper: _AggregativeBootstrap(ACC(Cls()), n_test_samples=1000, random_state=0), multiclass_method
|
|
#yield 'BootstrapEMQ', emq, on_calib_error='backup', val_split=5), emq_hyper, lambda hyper: _AggregativeBootstrap(EMQ(Cls(), on_calib_error='backup', calib=hyper['calib'], val_split=5), n_test_samples=1000, random_state=0), multiclass_method
|
|
#yield 'BootstrapHDy', hdy, hdy_hyper, lambda hyper: _AggregativeBootstrap(DMy(Cls(), **hyper), n_test_samples=1000, random_state=0), multiclass_method
|
|
#yield 'BootstrapKDEy', kde_gau, kdey_hyper, lambda hyper: _AggregativeBootstrap(KDEyML(Cls(), **hyper), n_test_samples=1000, random_state=0, verbose=True), multiclass_method
|
|
|
|
# Bayesian approaches:
|
|
# --------------------------------------------------------------------------------------------------------
|
|
# yield 'BayesianACC', acc, acc_hyper, lambda hyper: BayesianCC(Cls(), val_split=val_split, mcmc_seed=0), multiclass_method
|
|
#yield 'BayesianHDy', hdy, hdy_hyper, lambda hyper: PQ(Cls(), val_split=val_split, stan_seed=0, **hyper), only_binary
|
|
yield f'BaKDE-Ait-numpyro', kde_ait, kdey_hyper_clr, lambda hyper: BayesianKDEy(Cls(), kernel='aitchison', mcmc_seed=0, engine='numpyro', val_split=val_split, **hyper), multiclass_method
|
|
yield f'BaKDE-Gau-numpyro', kde_gau, kdey_hyper, lambda hyper: BayesianKDEy(Cls(), kernel='gaussian', mcmc_seed=0, engine='numpyro', val_split=val_split, **hyper), multiclass_method
|
|
yield f'BaKDE-Ait-T*', kde_ait, kdey_hyper_clr, lambda hyper: BayesianKDEy(Cls(),kernel='aitchison', mcmc_seed=0, engine='numpyro', temperature=None, val_split=val_split, **hyper), multiclass_method
|
|
yield f'BaKDE-Gau-T*', kde_gau, kdey_hyper, lambda hyper: BayesianKDEy(Cls(), kernel='gaussian', mcmc_seed=0, engine='numpyro', temperature=None, val_split=val_split, **hyper), multiclass_method
|
|
# yield 'BayEMQ', emq, acc_hyper, lambda hyper: BayesianMAPLS(Cls(), prior='uniform', temperature=1, exact_train_prev=False, val_split=val_split), multiclass_method
|
|
# yield 'BayEMQ*', emq, acc_hyper, lambda hyper: BayesianMAPLS(Cls(), prior='uniform', temperature=None, exact_train_prev=False, val_split=val_split), multiclass_method
|
|
|
|
|
|
def model_selection(dataset: DatasetHandler, point_quantifier: AggregativeQuantifier, grid: dict):
|
|
with qp.util.temp_seed(0):
|
|
print(f'performing model selection for {point_quantifier.__class__.__name__} with grid {grid}')
|
|
# model selection
|
|
if len(grid)>0:
|
|
train, val_prot = dataset.get_train_valprot_for_modsel()
|
|
mod_sel = GridSearchQ(
|
|
model=point_quantifier,
|
|
param_grid=grid,
|
|
protocol=val_prot,
|
|
refit=False,
|
|
n_jobs=-1,
|
|
verbose=True
|
|
).fit(*train.Xy)
|
|
best_params = mod_sel.best_params_
|
|
else:
|
|
best_params = {}
|
|
|
|
return best_params
|
|
|
|
|
|
def temperature_calibration(dataset: DatasetHandler, uncertainty_quantifier):
|
|
temperature = None
|
|
if hasattr(uncertainty_quantifier, 'temperature'):
|
|
if uncertainty_quantifier.temperature is None:
|
|
print('calibrating temperature')
|
|
train, val_prot = dataset.get_train_valprot_for_modsel()
|
|
if dataset.name.startswith('LeQua'):
|
|
temp_grid=[100., 500, 1000, 5_000, 10_000, 50_000]
|
|
else:
|
|
temp_grid=[.5, 1., 1.5, 2., 5., 10., 100.]
|
|
temperature = temp_calibration(uncertainty_quantifier, train, val_prot, temp_grid=temp_grid, n_jobs=-1)
|
|
uncertainty_quantifier.temperature = temperature
|
|
else:
|
|
temperature = uncertainty_quantifier.temperature
|
|
return temperature
|
|
|
|
|
|
|
|
def experiment(dataset: DatasetHandler, point_quantifier: AggregativeQuantifier, method_name:str, grid: dict, uncertainty_quant_constructor, hyper_choice_path: Path):
|
|
|
|
with qp.util.temp_seed(0):
|
|
|
|
# model selection
|
|
best_hyperparams = qp.util.pickled_resource(
|
|
hyper_choice_path, model_selection, dataset, cp(point_quantifier), grid
|
|
)
|
|
|
|
t_init = time()
|
|
uncertainty_quantifier = uncertainty_quant_constructor(best_hyperparams)
|
|
temperature = temperature_calibration(dataset, uncertainty_quantifier)
|
|
training, test_generator = dataset.get_train_testprot_for_eval()
|
|
uncertainty_quantifier.fit(*training.Xy)
|
|
tr_time = time() - t_init
|
|
|
|
# test
|
|
train_prevalence = training.prevalence()
|
|
results = defaultdict(list)
|
|
pbar = tqdm(enumerate(test_generator()), total=test_generator.total())
|
|
for i, (sample_X, true_prevalence) in pbar:
|
|
t_init = time()
|
|
point_estimate, region = uncertainty_quantifier.predict_conf(sample_X)
|
|
ttime = time()-t_init
|
|
|
|
results['true-prevs'].append(true_prevalence)
|
|
results['point-estim'].append(point_estimate)
|
|
results['shift'].append(qp.error.ae(true_prevalence, train_prevalence))
|
|
results['ae'].append(qp.error.ae(prevs_true=true_prevalence, prevs_hat=point_estimate))
|
|
results['rae'].append(qp.error.rae(prevs_true=true_prevalence, prevs_hat=point_estimate))
|
|
results['sre'].append(qp.error.sre(prevs_true=true_prevalence, prevs_hat=point_estimate, prevs_train=train_prevalence))
|
|
results['coverage'].append(region.coverage(true_prevalence))
|
|
results['amplitude'].append(region.montecarlo_proportion(n_trials=50_000))
|
|
results['test-time'].append(ttime)
|
|
results['samples'].append(region.samples)
|
|
|
|
pbar.set_description(f'{method_name} MAE={np.mean(results["ae"]):.5f} W={np.mean(results["sre"]):.5f} Cov={np.mean(results["coverage"]):.5f} AMP={np.mean(results["amplitude"]):.5f}')
|
|
|
|
report = {
|
|
'optim_hyper': best_hyperparams,
|
|
'train_time': tr_time,
|
|
'train-prev': train_prevalence,
|
|
'results': {k:np.asarray(v) for k,v in results.items()},
|
|
'temperature': temperature
|
|
}
|
|
|
|
return report
|
|
|
|
|
|
def check_skip_experiment(method_scope, dataset: DatasetHandler):
|
|
if method_scope == 'only_binary' and not dataset.is_binary():
|
|
return True
|
|
if method_scope == 'only_multiclass' and dataset.is_binary():
|
|
return True
|
|
return False
|
|
|
|
|
|
if __name__ == '__main__':
|
|
|
|
result_dir = RESULT_DIR
|
|
|
|
for data_handler in [MNISTHandler]:#, UCIMulticlassHandler,LeQuaHandler]:
|
|
for dataset in data_handler.iter():
|
|
qp.environ['SAMPLE_SIZE'] = dataset.sample_size
|
|
print(f'dataset={dataset.name}')
|
|
|
|
problem_type = 'binary' if dataset.is_binary() else 'multiclass'
|
|
|
|
for method_name, surrogate_quant, hyper_params, withconf_constructor, method_scope in methods(dataset):
|
|
if check_skip_experiment(method_scope, dataset):
|
|
continue
|
|
|
|
result_path = experiment_path(result_dir / problem_type, dataset.name, method_name)
|
|
hyper_path = experiment_path(result_dir / 'hyperparams' / problem_type, dataset.name, surrogate_quant.__class__.__name__)
|
|
|
|
report = qp.util.pickled_resource(
|
|
result_path, experiment, dataset, surrogate_quant, method_name, hyper_params, withconf_constructor, hyper_path
|
|
)
|
|
|
|
print(f'dataset={dataset.name}, '
|
|
f'method={method_name}: '
|
|
f'mae={report["results"]["ae"].mean():.5f}, '
|
|
f'W={report["results"]["sre"].mean():.5f}, '
|
|
f'coverage={report["results"]["coverage"].mean():.5f}, '
|
|
f'amplitude={report["results"]["amplitude"].mean():.5f}, ')
|
|
|
|
|
|
|
|
|