GridSearchQ adapted to work with generator functions and integrated for the baselines of LeQua2022; some tests with SVD

LeQua2022/TODO.txt

@@ -1,9 +1,13 @@
-1. los test hay que hacerlos suponiendo que las etiquetas no existen, es decir, viendo los resultados en los ficheros "prevalences" (renominar)
 2. tablas?
 3. fetch dataset (download, unzip, etc.)
 4. model selection
 5. plots
-6. estoy leyendo los samples en orden, y no hace falta. Sería mejor una función genérica que lee todos los ejemplos y
-    que de todos modos genera un output con el mismo nombre del file
-7. Make ResultSubmission class abstract, and create 4 instances thus forcing the field task_name to be set correctly
 8. No me convence que la lectura de los samples (caso en que no hay ground truth) viene en orden aleatorio
+9. Experimentar con vectores densos (PCA sobre tfidf por ejemplo)
+10. Si cambiamos el formato de los samples (por ejemplo, en lugar de svmlight con .txt a PCA con .dat) hay que cambiar
+    cosas en el código. Está escrito varias veces un glob(*.txt)
+11. Quitar las categorias como columnas de los ficheros de prevalences
+12. sample_size cannot be set to a non-integer in GridSearchQ whith protocol="gen" (it could, but is not indicated in doc)
+13. repair doc of GridSearchQ
+14. reparar la calibracion en LR (lo tuve que quitar para que funcionara GridSearchQ, y lo quité en todos los ficheros)
+15. podria poner que el eval_budget se usase en GridSearchQ con generator function para el progress bar de tqdm

LeQua2022/baselinesSVD_T1A.py

@@ -0,0 +1,84 @@
+import pickle
+
+import numpy as np
+from sklearn.linear_model import LogisticRegression
+from tqdm import tqdm
+import pandas as pd
+
+import quapy as qp
+from quapy.data import LabelledCollection
+from quapy.method.aggregative import *
+import quapy.functional as F
+from data import *
+import os
+import constants
+
+from sklearn.decomposition import TruncatedSVD
+
+
+# LeQua official baselines for task T1A (Binary/Vector)
+# =====================================================
+
+predictions_path = os.path.join('predictions', 'T1A')
+os.makedirs(predictions_path, exist_ok=True)
+
+models_path = os.path.join('models', 'T1A')
+os.makedirs(models_path, exist_ok=True)
+
+pathT1A = './data/T1A/public'
+T1A_devvectors_path = os.path.join(pathT1A, 'dev_vectors')
+T1A_devprevalence_path = os.path.join(pathT1A, 'dev_prevalences.csv')
+T1A_trainpath = os.path.join(pathT1A, 'training_vectors.txt')
+
+train = LabelledCollection.load(T1A_trainpath, load_binary_vectors)
+nF = train.instances.shape[1]
+svd = TruncatedSVD(n_components=300)
+train.instances = svd.fit_transform(train.instances)
+
+qp.environ['SAMPLE_SIZE'] = constants.T1A_SAMPLE_SIZE
+
+print(f'number of classes: {len(train.classes_)}')
+print(f'number of training documents: {len(train)}')
+print(f'training prevalence: {F.strprev(train.prevalence())}')
+print(f'training matrix shape: {train.instances.shape}')
+
+true_prevalence = ResultSubmission.load(T1A_devprevalence_path)
+
+for quantifier in [CC, ACC, PCC, PACC, EMQ, HDy]:
+
+    # classifier = CalibratedClassifierCV(LogisticRegression())
+    classifier = LogisticRegression()
+    model = quantifier(classifier).fit(train)
+    quantifier_name = model.__class__.__name__
+
+    predictions = ResultSubmission(categories=['negative', 'positive'])
+    for samplename, sample in tqdm(gen_load_samples_T1(T1A_devvectors_path, nF),
+                                   desc=quantifier_name, total=len(true_prevalence)):
+        sample = svd.transform(sample)
+        predictions.add(samplename, model.quantify(sample))
+
+    predictions.dump(os.path.join(predictions_path, quantifier_name + '.svd.csv'))
+    pickle.dump(model, open(os.path.join(models_path, quantifier_name+'.svd.pkl'), 'wb'), protocol=pickle.HIGHEST_PROTOCOL)
+
+    mae, mrae = evaluate_submission(true_prevalence, predictions)
+    print(f'{quantifier_name} mae={mae:.3f} mrae={mrae:.3f}')
+
+"""
+test:
+CC	0.1859	1.5406
+ACC	0.0453	0.2840
+PCC	0.1793	1.7187
+PACC	0.0287	0.1494
+EMQ	0.0225	0.1020
+HDy	0.0631	0.2307
+
+validation
+CC	0.1862	1.9587
+ACC	0.0394	0.2669
+PCC	0.1789	2.1383
+PACC	0.0354	0.1587
+EMQ	0.0224	0.0960
+HDy	0.0467	0.2121
+"""
+
+

LeQua2022/baselines_T1A.py

@@ -13,9 +13,16 @@ from data import *
 import os
 import constants
 
-predictions_path = os.path.join('predictions', 'T1A')  # binary - vector
+
+# LeQua official baselines for task T1A (Binary/Vector)
+# =====================================================
+
+predictions_path = os.path.join('predictions', 'T1A')
 os.makedirs(predictions_path, exist_ok=True)
 
+models_path = os.path.join('models', 'T1A')
+os.makedirs(models_path, exist_ok=True)
+
 pathT1A = './data/T1A/public'
 T1A_devvectors_path = os.path.join(pathT1A, 'dev_vectors')
 T1A_devprevalence_path = os.path.join(pathT1A, 'dev_prevalences.csv')
@@ -35,16 +42,19 @@ true_prevalence = ResultSubmission.load(T1A_devprevalence_path)
 
 for quantifier in [CC, ACC, PCC, PACC, EMQ, HDy]:
 
-    classifier = CalibratedClassifierCV(LogisticRegression())
+    # classifier = CalibratedClassifierCV(LogisticRegression(C=1))
+    classifier = LogisticRegression(C=1)
     model = quantifier(classifier).fit(train)
     quantifier_name = model.__class__.__name__
 
     predictions = ResultSubmission(categories=['negative', 'positive'])
-    for samplename, sample in tqdm(gen_load_samples_T1A(T1A_devvectors_path, nF),
+    for samplename, sample in tqdm(gen_load_samples_T1(T1A_devvectors_path, nF),
                                    desc=quantifier_name, total=len(true_prevalence)):
         predictions.add(samplename, model.quantify(sample))
 
     predictions.dump(os.path.join(predictions_path, quantifier_name + '.csv'))
+    pickle.dump(model, open(os.path.join(models_path, quantifier_name+'.pkl'), 'wb'), protocol=pickle.HIGHEST_PROTOCOL)
+
     mae, mrae = evaluate_submission(true_prevalence, predictions)
     print(f'{quantifier_name} mae={mae:.3f} mrae={mrae:.3f}')
 

LeQua2022/baselines_T1Amodsel.py

@@ -0,0 +1,91 @@
+import pickle
+
+import numpy as np
+from sklearn.linear_model import LogisticRegression
+from tqdm import tqdm
+import pandas as pd
+
+import quapy as qp
+from quapy.data import LabelledCollection
+from quapy.method.aggregative import *
+import quapy.functional as F
+from data import *
+import os
+import constants
+
+
+# LeQua official baselines for task T1A (Binary/Vector)
+# =====================================================
+
+predictions_path = os.path.join('predictions', 'T1A')
+os.makedirs(predictions_path, exist_ok=True)
+
+models_path = os.path.join('models', 'T1A')
+os.makedirs(models_path, exist_ok=True)
+
+pathT1A = './data/T1A/public'
+T1A_devvectors_path = os.path.join(pathT1A, 'dev_vectors')
+T1A_devprevalence_path = os.path.join(pathT1A, 'dev_prevalences.csv')
+T1A_trainpath = os.path.join(pathT1A, 'training_vectors.txt')
+
+train = LabelledCollection.load(T1A_trainpath, load_binary_vectors)
+nF = train.instances.shape[1]
+
+qp.environ['SAMPLE_SIZE'] = constants.T1A_SAMPLE_SIZE
+
+print(f'number of classes: {len(train.classes_)}')
+print(f'number of training documents: {len(train)}')
+print(f'training prevalence: {F.strprev(train.prevalence())}')
+print(f'training matrix shape: {train.instances.shape}')
+
+true_prevalence = ResultSubmission.load(T1A_devprevalence_path)
+
+param_grid = {'C': np.logspace(-3,3,7), 'class_weight': ['balanced', None]}
+
+
+def gen_samples():
+    return gen_load_samples_T1(T1A_devvectors_path, nF, ground_truth_path=T1A_devprevalence_path, return_filename=False)
+
+
+for quantifier in [CC, ACC, PCC, PACC, EMQ, HDy]:
+    #classifier = CalibratedClassifierCV(LogisticRegression(), n_jobs=-1)
+    classifier = LogisticRegression()
+    model = quantifier(classifier)
+    print(f'{model.__class__.__name__}: Model selection')
+    model = qp.model_selection.GridSearchQ(
+        model,
+        param_grid,
+        sample_size=None,
+        protocol='gen',
+        error=qp.error.mae,
+        refit=False,
+        verbose=True
+    ).fit(train, gen_samples)
+
+    quantifier_name = model.best_model().__class__.__name__
+    print(f'{quantifier_name} mae={model.best_score_:.3f} (params: {model.best_params_})')
+
+    pickle.dump(model.best_model(),
+                open(os.path.join(models_path, quantifier_name+'.modsel.pkl'), 'wb'),
+                protocol=pickle.HIGHEST_PROTOCOL)
+
+
+"""
+test:
+CC	0.1859	1.5406
+ACC	0.0453	0.2840
+PCC	0.1793	1.7187
+PACC	0.0287	0.1494
+EMQ	0.0225	0.1020
+HDy	0.0631	0.2307
+
+validation
+CC	0.1862	1.9587
+ACC	0.0394	0.2669
+PCC	0.1789	2.1383
+PACC	0.0354	0.1587
+EMQ	0.0224	0.0960
+HDy	0.0467	0.2121
+"""
+
+

LeQua2022/baselines_T1B.py

@@ -0,0 +1,55 @@
+import pickle
+
+import numpy as np
+from sklearn.linear_model import LogisticRegression
+from tqdm import tqdm
+import pandas as pd
+
+import quapy as qp
+from quapy.data import LabelledCollection
+from quapy.method.aggregative import *
+import quapy.functional as F
+from data import *
+import os
+import constants
+
+predictions_path = os.path.join('predictions', 'T1B')  # multiclass - vector
+os.makedirs(predictions_path, exist_ok=True)
+
+pathT1B = './data/T1B/public'
+T1B_devvectors_path = os.path.join(pathT1B, 'dev_vectors')
+T1B_devprevalence_path = os.path.join(pathT1B, 'dev_prevalences.csv')
+T1B_trainpath = os.path.join(pathT1B, 'training_vectors.txt')
+T1B_catmap = os.path.join(pathT1B, 'training_vectors_label_map.txt')
+
+train = LabelledCollection.load(T1B_trainpath, load_binary_vectors)
+nF = train.instances.shape[1]
+
+qp.environ['SAMPLE_SIZE'] = constants.T1B_SAMPLE_SIZE
+
+print(f'number of classes: {len(train.classes_)}')
+print(f'number of training documents: {len(train)}')
+print(f'training prevalence: {F.strprev(train.prevalence())}')
+print(f'training matrix shape: {train.instances.shape}')
+
+true_prevalence = ResultSubmission.load(T1B_devprevalence_path)
+
+cat2code, categories = load_category_map(T1B_catmap)
+
+for quantifier in [PACC]:  # [CC, ACC, PCC, PACC, EMQ]:
+
+    classifier = CalibratedClassifierCV(LogisticRegression())
+    model = quantifier(classifier).fit(train)
+    quantifier_name = model.__class__.__name__
+
+    predictions = ResultSubmission(categories=categories)
+    for samplename, sample in tqdm(gen_load_samples_T1(T1B_devvectors_path, nF),
+                                   desc=quantifier_name, total=len(true_prevalence)):
+        predictions.add(samplename, model.quantify(sample))
+
+    predictions.dump(os.path.join(predictions_path, quantifier_name + '.csv'))
+    mae, mrae = evaluate_submission(true_prevalence, predictions)
+    print(f'{quantifier_name} mae={mae:.3f} mrae={mrae:.3f}')
+
+
+

LeQua2022/constants.py

@@ -2,5 +2,6 @@ DEV_SAMPLES = 1000
 TEST_SAMPLES = 5000
 
 T1A_SAMPLE_SIZE = 250
+T1B_SAMPLE_SIZE = 1000
 
-ERROR_TOL=1E-3
+ERROR_TOL = 1E-3

LeQua2022/data.py

@@ -26,40 +26,45 @@ import constants
 def load_category_map(path):
     cat2code = {}
     with open(path, 'rt') as fin:
-        category, code = fin.readline().split()
+        for line in fin:
-        cat2code[category] = int(code)
+            category, code = line.split()
-    return cat2code
+            cat2code[category] = int(code)
+    code2cat = [cat for cat, code in sorted(cat2code.items(), key=lambda x:x[1])]
+    return cat2code, code2cat
 
 
 def load_binary_vectors(path, nF=None):
     return sklearn.datasets.load_svmlight_file(path, n_features=nF)
 
 
-def __gen_load_samples_with_groudtruth(path_dir:str, ground_truth_path:str, load_fn, **load_kwargs):
+def __gen_load_samples_with_groudtruth(path_dir:str, return_filename:bool, ground_truth_path:str, load_fn, **load_kwargs):
     true_prevs = ResultSubmission.load(ground_truth_path)
     for filename, prevalence in true_prevs.iterrows():
         sample, _ = load_fn(os.path.join(path_dir, filename), **load_kwargs)
-        yield filename, sample, prevalence
+        if return_filename:
+            yield filename, sample, prevalence
+        else:
+            yield sample, prevalence
 
 
-def __gen_load_samples_without_groudtruth(path_dir:str, load_fn, **load_kwargs):
+def __gen_load_samples_without_groudtruth(path_dir:str, return_filename:bool, load_fn, **load_kwargs):
     for filepath in glob(os.path.join(path_dir, '*_sample_*.txt')):
         sample, _ = load_fn(filepath, **load_kwargs)
-        yield os.path.basename(filepath), sample
+        if return_filename:
+            yield os.path.basename(filepath), sample
+        else:
+            yield sample
 
 
-def gen_load_samples_T1A(path_dir:str, nF:int, ground_truth_path:str = None):
+def gen_load_samples_T1(path_dir:str, nF:int, ground_truth_path:str = None, return_filename=True):
     if ground_truth_path is None:
-        for filename, sample in __gen_load_samples_without_groudtruth(path_dir, load_binary_vectors, nF=nF):
+        # the generator function returns tuples (filename:str, sample:csr_matrix)
-            yield filename, sample
+        gen_fn = __gen_load_samples_without_groudtruth(path_dir, return_filename, load_binary_vectors, nF=nF)
     else:
-        for filename, sample, prevalence in __gen_load_samples_with_groudtruth(path_dir, ground_truth_path, load_binary_vectors, nF=nF):
+        # the generator function returns tuples (filename:str, sample:csr_matrix, prevalence:ndarray)
-            yield filename, sample, prevalence
+        gen_fn = __gen_load_samples_with_groudtruth(path_dir, return_filename, ground_truth_path, load_binary_vectors, nF=nF)
-
+    for r in gen_fn:
-
+        yield r
-def gen_load_samples_T1B(path_dir:str, ground_truth_path:str = None):
-    # for ... : yield
-    pass
 
 
 def gen_load_samples_T2A(path_dir:str, ground_truth_path:str = None):

quapy/evaluation.py

@@ -9,6 +9,7 @@ from quapy.method.base import BaseQuantifier
 from quapy.util import temp_seed
 import quapy.functional as F
 import pandas as pd
+import inspect
 
 
 def artificial_prevalence_prediction(
@@ -78,6 +79,27 @@ def natural_prevalence_prediction(
     return _predict_from_indexes(indexes, model, test, n_jobs, verbose)
 
 
+def gen_prevalence_prediction(model: BaseQuantifier, gen_fn: Callable, eval_budget=None):
+    if not inspect.isgenerator(gen_fn()):
+        raise ValueError('param "gen_fun" is not a generator')
+
+    if not isinstance(eval_budget, int):
+        eval_budget = -1
+
+    true_prevalences, estim_prevalences = [], []
+    for sample_instances, true_prev in gen_fn():
+        true_prevalences.append(true_prev)
+        estim_prevalences.append(model.quantify(sample_instances))
+        eval_budget -= 1
+        if eval_budget == 0:
+            break
+
+    true_prevalences = np.asarray(true_prevalences)
+    estim_prevalences = np.asarray(estim_prevalences)
+
+    return true_prevalences, estim_prevalences
+
+
 def _predict_from_indexes(
         indexes,
         model: BaseQuantifier,

quapy/model_selection.py

@@ -5,8 +5,9 @@ from typing import Union, Callable
 
 import quapy as qp
 from quapy.data.base import LabelledCollection
-from quapy.evaluation import artificial_prevalence_prediction, natural_prevalence_prediction
+from quapy.evaluation import artificial_prevalence_prediction, natural_prevalence_prediction, gen_prevalence_prediction
 from quapy.method.aggregative import BaseQuantifier
+import inspect
 
 
 class GridSearchQ(BaseQuantifier):
@@ -74,8 +75,10 @@ class GridSearchQ(BaseQuantifier):
         self.timeout = timeout
         self.verbose = verbose
         self.__check_error(error)
-        assert self.protocol in {'app', 'npp'}, \
+        assert self.protocol in {'app', 'npp', 'gen'}, \
-            'unknown protocol; valid ones are "app" or "npp" for the "artificial" or the "natural" prevalence protocols'
+            'unknown protocol: valid ones are "app" or "npp" for the "artificial" or the "natural" prevalence ' \
+            'protocols. Use protocol="gen" when passing a generator function thorough val_split that yields a ' \
+            'sample (instances) and their prevalence (ndarray) at each iteration.'
         if self.protocol == 'npp':
             if self.n_repetitions is None or self.n_repetitions == 1:
                 if self.eval_budget is not None:
@@ -99,9 +102,14 @@ class GridSearchQ(BaseQuantifier):
             assert 0. < validation < 1., 'validation proportion should be in (0,1)'
             training, validation = training.split_stratified(train_prop=1 - validation)
             return training, validation
+        elif self.protocol=='gen' and inspect.isgenerator(validation()):
+            return training, validation
         else:
             raise ValueError(f'"validation" must either be a LabelledCollection or a float in (0,1) indicating the'
-                             f'proportion of training documents to extract (type found: {type(validation)})')
+                             f'proportion of training documents to extract (type found: {type(validation)}). '
+                             f'Optionally, "validation" can be a callable function returning a generator that yields '
+                             f'the sample instances along with their true prevalence at each iteration by '
+                             f'setting protocol="gen".')
 
     def __check_error(self, error):
         if error in qp.error.QUANTIFICATION_ERROR:
@@ -132,6 +140,8 @@ class GridSearchQ(BaseQuantifier):
             return natural_prevalence_prediction(
                 model, val_split, self.sample_size,
                 **commons)
+        elif self.protocol == 'gen':
+            return gen_prevalence_prediction(model, gen_fn=val_split, eval_budget=self.eval_budget)
         else:
             raise ValueError('unknown protocol')
 
@@ -144,7 +154,8 @@ class GridSearchQ(BaseQuantifier):
         if val_split is None:
             val_split = self.val_split
         training, val_split = self.__check_training_validation(training, val_split)
-        assert isinstance(self.sample_size, int) and self.sample_size > 0, 'sample_size must be a positive integer'
+        if self.protocol != 'gen':
+            assert isinstance(self.sample_size, int) and self.sample_size > 0, 'sample_size must be a positive integer'
 
         params_keys = list(self.param_grid.keys())
         params_values = list(self.param_grid.values())
@@ -192,8 +203,6 @@ class GridSearchQ(BaseQuantifier):
             raise TimeoutError('all jobs took more than the timeout time to end')
 
         self.sout(f'optimization finished: best params {self.best_params_} (score={self.best_score_:.5f})')
-        # model.set_params(**self.best_params_)
-        # self.best_model_ = deepcopy(model)
 
         if self.refit:
             self.sout(f'refitting on the whole development set')
@@ -203,11 +212,11 @@ class GridSearchQ(BaseQuantifier):
 
     def quantify(self, instances):
         assert hasattr(self, 'best_model_'), 'quantify called before fit'
-        return self.best_model_.quantify(instances)
+        return self.best_model().quantify(instances)
 
     @property
     def classes_(self):
-        return self.best_model_.classes_
+        return self.best_model().classes_
 
     def set_params(self, **parameters):
         self.param_grid = parameters