adapting new format

LeQua2022/baselines_T1.py

@@ -14,10 +14,10 @@ import constants
 
 def baselines():
     yield CC(LR(n_jobs=-1)), "CC"
-    yield ACC(LR(n_jobs=-1)), "ACC"
+    # yield ACC(LR(n_jobs=-1)), "ACC"
-    yield PCC(LR(n_jobs=-1)), "PCC"
+    # yield PCC(LR(n_jobs=-1)), "PCC"
-    yield PACC(LR(n_jobs=-1)), "PACC"
+    # yield PACC(LR(n_jobs=-1)), "PACC"
-    yield EMQ(CalibratedClassifierCV(LR(), n_jobs=-1)), "SLD"
+    # yield EMQ(CalibratedClassifierCV(LR(), n_jobs=-1)), "SLD"
     # yield HDy(LR(n_jobs=-1)) if args.task == 'T1A' else OneVsAll(HDy(LR()), n_jobs=-1), "HDy"
     # yield MLPE(), "MLPE"
 
@@ -28,7 +28,7 @@ def main(args):
 
     path_dev_vectors = os.path.join(args.datadir, 'dev_vectors')
     path_dev_prevs = os.path.join(args.datadir, 'dev_prevalences.csv')
-    path_train = os.path.join(args.datadir, 'training_vectors.txt')
+    path_train = os.path.join(args.datadir, 'training_vectors.csv')
 
     qp.environ['SAMPLE_SIZE'] = constants.SAMPLE_SIZE[args.task]
 
@@ -46,13 +46,15 @@ def main(args):
     # }
 
     param_grid = {
-        'C': [1],
+        'C': [0.01],
         'class_weight': ['balanced']
     }
+    target_metric = qp.error.mrae
 
     def gen_samples():
         return gen_load_samples(path_dev_vectors, ground_truth_path=path_dev_prevs, return_id=False,
-                                load_fn=load_vector_documents, nF=nF)
+                                load_fn=load_vector_documents, ext='csv')
+
 
     for quantifier, q_name in baselines():
         print(f'{q_name}: Model selection')
@@ -61,12 +63,12 @@ def main(args):
             param_grid,
             sample_size=None,
             protocol='gen',
-            error=qp.error.mae,
+            error=target_metric,  #qp.error.mae,
             refit=False,
             verbose=True
         ).fit(train, gen_samples)
 
-        print(f'{q_name} got MAE={quantifier.best_score_:.3f} (hyper-params: {quantifier.best_params_})')
+        print(f'{q_name} got MRAE={quantifier.best_score_:.5f} (hyper-params: {quantifier.best_params_})')
 
         model_path = os.path.join(models_path, q_name+'.pkl')
         print(f'saving model in {model_path}')
@@ -91,8 +93,8 @@ if __name__ == '__main__':
         raise ValueError(f'path {args.datadir} is not a valid directory')
     if not os.path.exists(os.path.join(args.datadir, "dev_prevalences.csv")):
         raise FileNotFoundError(f'path {args.datadir} does not contain "dev_prevalences.csv" file')
-    if not os.path.exists(os.path.join(args.datadir, "training_vectors.txt")):
+    if not os.path.exists(os.path.join(args.datadir, "training_vectors.csv")):
-        raise FileNotFoundError(f'path {args.datadir} does not contain "training_vectors.txt" file')
+        raise FileNotFoundError(f'path {args.datadir} does not contain "training_vectors.csv" file')
     if not os.path.exists(os.path.join(args.datadir, "dev_vectors")):
         raise FileNotFoundError(f'path {args.datadir} does not contain "dev_vectors" folder')
 

LeQua2022/baselines_T2.py

@@ -1,8 +1,14 @@
 import argparse
 import pickle
 
+from sklearn.decomposition import TruncatedSVD
 from sklearn.feature_extraction.text import TfidfVectorizer
 from sklearn.linear_model import LogisticRegression as LR
+from sklearn.pipeline import Pipeline
+from sklearn.preprocessing import StandardScaler
+
+from LeQua2022.pretrained_embeddings import TfidfWordEmbeddingTransformer, WordEmbeddingAverageTransformer
+from LeQua2022.word_class_embeddings import WordClassEmbeddingsTransformer, ConcatenateEmbeddingsTransformer
 from quapy.method.aggregative import *
 from quapy.method.non_aggregative import MaximumLikelihoodPrevalenceEstimation as MLPE
 import quapy.functional as F
@@ -20,7 +26,7 @@ def baselines():
     yield PCC(LR(n_jobs=-1)), "PCC"
     yield PACC(LR(n_jobs=-1)), "PACC"
     yield EMQ(CalibratedClassifierCV(LR(), n_jobs=-1)), "SLD"
-    yield HDy(LR(n_jobs=-1)) if args.task == 'T2A' else OneVsAll(HDy(LR()), n_jobs=-1), "HDy"
+    # yield HDy(LR(n_jobs=-1)) if args.task == 'T2A' else OneVsAll(HDy(LR()), n_jobs=-1), "HDy"
     # yield MLPE(), "MLPE"
 
 
@@ -35,9 +41,69 @@ def main(args):
     qp.environ['SAMPLE_SIZE'] = constants.SAMPLE_SIZE[args.task]
 
     train = LabelledCollection.load(path_train, load_raw_documents)
-    tfidf = TfidfVectorizer(lowercase=True, stop_words='english', min_df=4)  # TfidfVectorizer(min_df=5)
+
-    train.instances = tfidf.fit_transform(train.instances)
+    if args.mode == 'tfidf1':
-    nF = train.instances.shape[1]
+        tfidf = TfidfVectorizer(min_df=5, sublinear_tf=True)
+    if args.mode == 'tfidf2':
+        tfidf = TfidfVectorizer(min_df=5, sublinear_tf=True, ngram_range=(1,2))
+    if args.mode == 'tfidf3':
+        tfidf = Pipeline([
+            ('tfidf', TfidfVectorizer(min_df=5, sublinear_tf=True)),
+            ('svd', TruncatedSVD(n_components=300))
+        ])
+    if args.mode == 'tfidf4':
+        tfidf = Pipeline([
+            ('tfidf', TfidfVectorizer(min_df=5, sublinear_tf=True, ngram_range=(1,2))),
+            ('svd', TruncatedSVD(n_components=300))
+        ])
+    if args.mode == 'glove1':
+        tfidf = Pipeline([
+            ('glove-ave', WordEmbeddingAverageTransformer(wordset_name='glove', path='/mnt/1T/Datasets/GloVe')),
+            ('zscore', StandardScaler())
+        ])
+    if args.mode == 'glove2':
+        tfidf = WordEmbeddingAverageTransformer(wordset_name='glove', path='/mnt/1T/Datasets/GloVe')
+    if args.mode == 'glove3':
+        vect = TfidfVectorizer(min_df=5, sublinear_tf=True)
+        tfidf = Pipeline([
+            ('tfidf', vect),
+            ('embedding', TfidfWordEmbeddingTransformer(
+                wordset_name='glove',
+                features_call=vect.get_feature_names_out,
+                path='/mnt/1T/Datasets/GloVe')),
+            ('zscore', StandardScaler())
+        ])
+    if args.mode == 'glove4':
+        vect = TfidfVectorizer(min_df=5, sublinear_tf=True)
+        tfidf = Pipeline([
+            ('tfidf', vect),
+            ('embedding', TfidfWordEmbeddingTransformer(
+                wordset_name='glove',
+                features_call=vect.get_feature_names_out,
+                path='/mnt/1T/Datasets/GloVe'))
+        ])
+    if args.mode == 'wce1':
+        tfidf = WordClassEmbeddingsTransformer()
+    if args.mode == 'wce2':
+        glove = Pipeline([
+            ('glove-ave', WordEmbeddingAverageTransformer(wordset_name='glove', path='/mnt/1T/Datasets/GloVe')),
+            ('zscore', StandardScaler())
+        ])
+        wce = WordClassEmbeddingsTransformer()
+        tfidf = ConcatenateEmbeddingsTransformer([glove, wce])
+    if args.mode == 'wce3':
+        glove = Pipeline([
+            ('glove-ave', WordEmbeddingAverageTransformer(wordset_name='glove', path='/mnt/1T/Datasets/GloVe')),
+            ('zscore', StandardScaler())
+        ])
+        wce = WordClassEmbeddingsTransformer()
+        tfidf = Pipeline([
+            ('glove-wce', ConcatenateEmbeddingsTransformer([glove, wce])),
+            ('svd', TruncatedSVD(n_components=300))
+            ])
+    target_metric = qp.error.mrae
+
+    train.instances = tfidf.fit_transform(*train.Xy)
 
     print(f'number of classes: {len(train.classes_)}')
     print(f'number of training documents: {len(train)}')
@@ -58,6 +124,7 @@ def main(args):
         return gen_load_samples(path_dev_vectors, ground_truth_path=path_dev_prevs, return_id=False,
                                 load_fn=load_raw_unlabelled_documents, vectorizer=tfidf)
 
+    outs = []
     for quantifier, q_name in baselines():
         print(f'{q_name}: Model selection')
         quantifier = qp.model_selection.GridSearchQ(
@@ -65,17 +132,25 @@ def main(args):
             param_grid,
             sample_size=None,
             protocol='gen',
-            error=qp.error.mae,
+            error=target_metric,  #qp.error.mae,
             refit=False,
             verbose=True
         ).fit(train, gen_samples)
 
-        print(f'{q_name} got MAE={quantifier.best_score_:.3f} (hyper-params: {quantifier.best_params_})')
+        print(f'{q_name} got MAE={quantifier.best_score_:.5f} (hyper-params: {quantifier.best_params_})')
+        outs.append(f'{q_name} got MAE={quantifier.best_score_:.5f} (hyper-params: {quantifier.best_params_})')
 
         model_path = os.path.join(models_path, q_name+'.'+args.task+'.pkl')
         print(f'saving model in {model_path}')
         pickle.dump(quantifier.best_model(), open(model_path, 'wb'), protocol=pickle.HIGHEST_PROTOCOL)
 
+    print(tfidf)
+    print(args.mode)
+    print(outs)
+    with open(f'{args.mode}.{args.task}.txt', 'wt') as foo:
+        for line in outs:
+            foo.write(f'{line}\n')
+
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser(description='LeQua2022 Task T2A/T2B baselines')
@@ -87,6 +162,8 @@ if __name__ == '__main__':
     parser.add_argument('modeldir', metavar='MODEL-PATH', type=str,
                         help='Path where to save the models. '
                              'A subdirectory named <task> will be automatically created.')
+    parser.add_argument('mode', metavar='PREPROCESSMODE', type=str,
+                        help='modality of preprocessing')
     args = parser.parse_args()
 
     if not os.path.exists(args.datadir):

LeQua2022/data.py

@@ -34,33 +34,42 @@ def load_raw_unlabelled_documents(path, vectorizer=None):
     return documents, None
 
 
-def load_vector_documents(path, nF=None):
+# def load_vector_documents(path, nF=None):
-    X, y = sklearn.datasets.load_svmlight_file(path, n_features=nF)
+#     X, y = sklearn.datasets.load_svmlight_file(path, n_features=nF, zero_based=True)
-    y = y.astype(int)
+#     y = y.astype(int)
+#     return X, y
+
+def load_vector_documents(path):
+    D = pd.read_csv(path).to_numpy(dtype=np.float)
+    labelled = D.shape[1] == 301
+    if labelled:
+        X, y = D[:,:300], D[:,-1].astype(np.int).flatten()
+    else:
+        X, y = D, None
     return X, y
 
 
-def __gen_load_samples_with_groudtruth(path_dir:str, return_id:bool, ground_truth_path:str, load_fn, **load_kwargs):
+def __gen_load_samples_with_groudtruth(path_dir:str, return_id:bool, ground_truth_path:str, ext:str, load_fn, **load_kwargs):
     true_prevs = ResultSubmission.load(ground_truth_path)
     for id, prevalence in true_prevs.iterrows():
-        sample, _ = load_fn(os.path.join(path_dir, f'{id}.txt'), **load_kwargs)
+        sample, _ = load_fn(os.path.join(path_dir, f'{id}.{ext}'), **load_kwargs)
         yield (id, sample, prevalence) if return_id else (sample, prevalence)
 
 
-def __gen_load_samples_without_groudtruth(path_dir:str, return_id:bool, load_fn, **load_kwargs):
+def __gen_load_samples_without_groudtruth(path_dir:str, return_id:bool, ext:str, load_fn, **load_kwargs):
-    nsamples = len(glob(os.path.join(path_dir, '*.txt')))
+    nsamples = len(glob(os.path.join(path_dir, f'*.{ext}')))
     for id in range(nsamples):
-        sample, _ = load_fn(os.path.join(path_dir, f'{id}.txt'), **load_kwargs)
+        sample, _ = load_fn(os.path.join(path_dir, f'{id}.{ext}'), **load_kwargs)
         yield (id, sample) if return_id else sample
 
 
-def gen_load_samples(path_dir:str, ground_truth_path:str = None, return_id=True, load_fn=load_vector_documents, **load_kwargs):
+def gen_load_samples(path_dir:str, ground_truth_path:str = None, return_id=True, ext='txt', load_fn=load_vector_documents, **load_kwargs):
     if ground_truth_path is None:
         # the generator function returns tuples (docid:str, sample:csr_matrix or str)
-        gen_fn = __gen_load_samples_without_groudtruth(path_dir, return_id, load_fn, **load_kwargs)
+        gen_fn = __gen_load_samples_without_groudtruth(path_dir, return_id, ext, load_fn, **load_kwargs)
     else:
         # the generator function returns tuples (docid:str, sample:csr_matrix or str, prevalence:ndarray)
-        gen_fn = __gen_load_samples_with_groudtruth(path_dir, return_id, ground_truth_path, load_fn, **load_kwargs)
+        gen_fn = __gen_load_samples_with_groudtruth(path_dir, return_id, ground_truth_path, ext, load_fn, **load_kwargs)
     for r in gen_fn:
         yield r
 
@@ -139,7 +148,11 @@ class ResultSubmission:
 
     @classmethod
     def check_file_format(cls, path) -> Union[pd.DataFrame, Tuple[pd.DataFrame, str]]:
-        df = pd.read_csv(path, index_col=0)
+        try:
+            df = pd.read_csv(path, index_col=0)
+        except Exception as e:
+            print(f'the file {path} does not seem to be a valid csv file. ')
+            print(e)
         return ResultSubmission.check_dataframe_format(df, path=path)
 
     @classmethod

quapy/evaluation.py

@@ -24,6 +24,7 @@ def artificial_prevalence_prediction(
         verbose=False):
     """
     Performs the predictions for all samples generated according to the artificial sampling protocol.
+
     :param model: the model in charge of generating the class prevalence estimations
     :param test: the test set on which to perform arificial sampling
     :param sample_size: the size of the samples

quapy/model_selection.py

@@ -3,6 +3,8 @@ import signal
 from copy import deepcopy
 from typing import Union, Callable
 
+import numpy as np
+
 import quapy as qp
 from quapy.data.base import LabelledCollection
 from quapy.evaluation import artificial_prevalence_prediction, natural_prevalence_prediction, gen_prevalence_prediction
@@ -190,6 +192,7 @@ class GridSearchQ(BaseQuantifier):
                 model.fit(training)
                 true_prevalences, estim_prevalences = self.__generate_predictions(model, val_split)
                 score = self.error(true_prevalences, estim_prevalences)
+
                 self._sout(f'checking hyperparams={params} got {self.error.__name__} score {score:.5f}')
                 if self.best_score_ is None or score < self.best_score_:
                     self.best_score_ = score