import fix

BayesianKDEy/_bayeisan_kdey.py

@@ -43,6 +43,7 @@ class BayesianKDEy(AggregativeSoftQuantifier, KDEBase, WithConfidenceABC):
                  region: str = 'intervals',
                  explore='simplex',
                  step_size=0.05,
+                 temperature=1.,
                  verbose: bool = False):
 
         if num_warmup <= 0:
@@ -51,6 +52,7 @@ class BayesianKDEy(AggregativeSoftQuantifier, KDEBase, WithConfidenceABC):
             raise ValueError(f'parameter {num_samples=} must be a positive integer')
         assert explore in ['simplex', 'clr', 'ilr'], \
             f'unexpected value for param {explore=}; valid ones are "simplex", "clr", and "ilr"'
+        assert temperature>0., f'temperature must be >0'
 
         super().__init__(classifier, fit_classifier, val_split)
         self.bandwidth = KDEBase._check_bandwidth(bandwidth, kernel)
@@ -62,6 +64,7 @@ class BayesianKDEy(AggregativeSoftQuantifier, KDEBase, WithConfidenceABC):
         self.region = region
         self.explore = explore
         self.step_size = step_size
+        self.temperature = temperature
         self.verbose = verbose
 
     def aggregation_fit(self, classif_predictions, labels):
@@ -99,7 +102,7 @@ class BayesianKDEy(AggregativeSoftQuantifier, KDEBase, WithConfidenceABC):
         def log_likelihood(prev, epsilon=1e-10):
             test_likelihoods = prev @ test_densities
             test_loglikelihood = np.log(test_likelihoods + epsilon)
-            return np.sum(test_loglikelihood)
+            return (1./self.temperature) * np.sum(test_loglikelihood)
 
         # def log_prior(prev):
             # todo: adapt to arbitrary prior knowledge (e.g., something around training prevalence)
@@ -167,7 +170,8 @@ class BayesianKDEy(AggregativeSoftQuantifier, KDEBase, WithConfidenceABC):
                 recent_accept_rate = np.mean(acceptance_history[-100:])
                 step_size *= np.exp(adapt_rate * (recent_accept_rate - target_acceptance))
                 # step_size = float(np.clip(step_size, min_step, max_step))
-                # print(f'acceptance-rate={recent_accept_rate*100:.3f}%, step-size={step_size:.5f}')
+                if i %100==0:
+                    print(f'acceptance-rate={recent_accept_rate*100:.3f}%, step-size={step_size:.5f}')
 
         # remove "warmup" initial iterations
         samples = np.asarray(samples[self.num_warmup:])

BayesianKDEy/single_experiment_debug.py

@@ -29,7 +29,7 @@ from time import time
 from sklearn.base import clone, BaseEstimator
 
 
-def method():
+def methods():
     """
     Returns a tuple (name, quantifier, hyperparams, bayesian/bootstrap_constructor), where:
     - name: is a str representing the name of the method (e.g., 'BayesianKDEy')
@@ -47,40 +47,33 @@ def method():
 
     # yield 'BootstrapKDEy', KDEyML(LR()), kdey_hyper, lambda hyper: AggregativeBootstrap(KDEyML(LR(), **hyper), n_test_samples=1000, random_state=0, verbose=True),
     # yield 'BayesianKDEy', KDEyML(LR()), kdey_hyper, lambda hyper: BayesianKDEy(mcmc_seed=0, **hyper),
-    return 'BayKDE*CLR', KDEyCLR(LR()), kdey_hyper_clr, lambda hyper: BayesianKDEy(kernel='aitchison', mcmc_seed=0,
-                                                                                   explore=True,
-                                                                                   step_size=.15,
-                                                                                   # num_warmup = 5000,
-                                                                                   # num_samples = 10_000,
-                                                                                   # region='ellipse',
-                                                                                   **hyper),
+    for T in [1., 10., 100., 1000.]:
+        yield (f'BaKDE-CLR-T{T}', KDEyCLR(LR()), kdey_hyper_clr,
+                lambda hyper: BayesianKDEy(kernel='aitchison', mcmc_seed=0, temperature=T, step_size=.15, **hyper)),
 
 
 
 if __name__ == '__main__':
 
     binary = {
-        'datasets': qp.datasets.UCI_BINARY_DATASETS,
         'fetch_fn': qp.datasets.fetch_UCIBinaryDataset,
         'sample_size': 500
     }
 
     multiclass = {
-        'datasets': qp.datasets.UCI_MULTICLASS_DATASETS,
         'fetch_fn': qp.datasets.fetch_UCIMulticlassDataset,
         'sample_size': 1000
     }
 
-    result_dir = Path('./results')
-
     setup = multiclass
-    qp.environ['SAMPLE_SIZE'] = setup['sample_size']
     data_name = 'digits'
+
+    qp.environ['SAMPLE_SIZE'] = setup['sample_size']
     print(f'dataset={data_name}')
     data = setup['fetch_fn'](data_name)
     is_binary = data.n_classes==2
-    hyper_subdir  = result_dir / 'hyperparams' / ('binary' if is_binary else 'multiclass')
-    method_name, method, hyper_params, withconf_constructor = method()
+    hyper_subdir  = Path('./results') / 'hyperparams' / ('binary' if is_binary else 'multiclass')
+    for method_name, method, hyper_params, withconf_constructor in methods():
         hyper_path = experiment_path(hyper_subdir, data_name, method.__class__.__name__)
         report = experiment(data, method, method_name, hyper_params, withconf_constructor, hyper_path)
 

quapy/method/confidence.py

@@ -5,7 +5,7 @@ from sklearn.metrics import confusion_matrix
 
 import quapy as qp
 import quapy.functional as F
-from functional import CompositionalTransformation, CLRtransformation, ILRtransformation
+from quapy.functional import CompositionalTransformation, CLRtransformation, ILRtransformation
 from quapy.method import _bayesian
 from quapy.data import LabelledCollection
 from quapy.method.aggregative import AggregativeQuantifier, AggregativeCrispQuantifier, AggregativeSoftQuantifier, BinaryAggregativeQuantifier