Improve temporary data management using temp directories

delft/applications/citationClassifier.py

@@ -5,6 +5,7 @@
 import argparse
 import time
 from delft.textClassification.models import architectures
+from delft.utilities.misc import DEFAULT_DATA_MODEL_PATH_TEXT_CLASSIFICATION
 
 list_classes = [
                     "negative", 
@@ -34,7 +35,7 @@ def configure(architecture):
     return batch_size, maxlen, patience, early_stop, max_epoch
 
 
-def train(embeddings_name, fold_count, architecture="gru", transformer=None):
+def train(embeddings_name, fold_count, architecture="gru", transformer=None, output_directory=None):
     batch_size, maxlen, patience, early_stop, max_epoch = configure(architecture)
 
     model = Classifier('citations_'+architecture, architecture=architecture, list_classes=list_classes, max_epoch=max_epoch, fold_number=fold_count, 
@@ -48,11 +49,12 @@ def train(embeddings_name, fold_count, architecture="gru", transformer=None):
         model.train(xtr, y)
     else:
         model.train_nfold(xtr, y)
+
     # saving the model
-    model.save()
+    model.save(output_directory)
 
 
-def train_and_eval(embeddings_name, fold_count, architecture="gru", transformer=None): 
+def train_and_eval(embeddings_name, fold_count, architecture="gru", transformer=None, output_directory=None):
     batch_size, maxlen, patience, early_stop, max_epoch = configure(architecture)
 
     model = Classifier('citations_'+architecture, architecture=architecture, list_classes=list_classes, max_epoch=max_epoch, fold_number=fold_count, 
@@ -69,9 +71,8 @@ def train_and_eval(embeddings_name, fold_count, architecture="gru", transformer=
         model.train(x_train, y_train)
     else:
         model.train_nfold(x_train, y_train)
-
-    # saving the model
-    model.save()
+
+    model.save(output_directory)
 
     model.eval(x_test, y_test)
 
@@ -118,6 +119,7 @@ def classify(texts, output_format, architecture="gru", embeddings_name=None, tra
             "HuggingFace transformers hub will be used otherwise to fetch the model, see https://huggingface.co/models " + \
             "for model names"
     )
+    parser.add_argument("--output", help="Directory where to save a trained model.", default=DEFAULT_DATA_MODEL_PATH_TEXT_CLASSIFICATION)
 
     args = parser.parse_args()
 
@@ -126,26 +128,27 @@ def classify(texts, output_format, architecture="gru", embeddings_name=None, tra
 
     embeddings_name = args.embedding
     transformer = args.transformer
+    output = args.output
 
     architecture = args.architecture
     if architecture not in architectures:
         print('unknown model architecture, must be one of '+str(architectures))
 
-    if transformer == None and embeddings_name == None:
+    if transformer is None and embeddings_name is None:
         # default word embeddings
         embeddings_name = "glove-840B"
 
     if args.action == 'train':
         if args.fold_count < 1:
             raise ValueError("fold-count should be equal or more than 1")
 
-        train(embeddings_name, args.fold_count, architecture=architecture, transformer=transformer)
+        train(embeddings_name, args.fold_count, architecture=architecture, transformer=transformer, output_directory=output)
 
     if args.action == 'train_eval':
         if args.fold_count < 1:
             raise ValueError("fold-count should be equal or more than 1")
 
-        y_test = train_and_eval(embeddings_name, args.fold_count, architecture=architecture, transformer=transformer)    
+        y_test = train_and_eval(embeddings_name, args.fold_count, architecture=architecture, transformer=transformer, output_directory=output)
 
     if args.action == 'classify':
         someTexts = ['One successful strategy [15] computes the set-similarity involving (multi-word) keyphrases about the mentions and the entities, collected from the KG.', 

delft/applications/dataseerClassifier.py

@@ -1,15 +1,15 @@
 import json
-from delft.utilities.Embeddings import Embeddings
 from delft.utilities.Utilities import split_data_and_labels
 from delft.textClassification.reader import load_dataseer_corpus_csv
 from delft.textClassification.reader import vectorize as vectorizer
-import delft.textClassification
 from delft.textClassification import Classifier
 import argparse
 import time
 from delft.textClassification.models import architectures
 import numpy as np
 
+from delft.utilities.misc import DEFAULT_DATA_MODEL_PATH_TEXT_CLASSIFICATION
+
 """
     Classifier for deciding if a sentence introduce a dataset or not, and prediction of the 
     dataset type. 
@@ -33,7 +33,7 @@ def configure(architecture):
     return batch_size, maxlen, patience, early_stop, max_epoch
 
 
-def train(embeddings_name, fold_count, architecture="gru", transformer=None, cascaded=False): 
+def train(embeddings_name, fold_count, architecture="gru", transformer=None, cascaded=False, output_directory=None):
     print('loading binary dataset type corpus...')
     xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-binary.csv")
 
@@ -50,8 +50,8 @@ def train(embeddings_name, fold_count, architecture="gru", transformer=None, cas
         model.train(xtr, y)
     else:
         model.train_nfold(xtr, y)
-    # saving the model
-    model.save()
+
+    model.save(output_directory)
 
     print('loading reuse dataset type corpus...')
     xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-reuse.csv")
@@ -67,8 +67,8 @@ def train(embeddings_name, fold_count, architecture="gru", transformer=None, cas
         model.train(xtr, y)
     else:
         model.train_nfold(xtr, y)
-    # saving the model
-    model.save()
+
+    model.save(output_directory)
 
     print('loading first-level dataset type corpus...')
     xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-1.csv")
@@ -85,9 +85,10 @@ def train(embeddings_name, fold_count, architecture="gru", transformer=None, cas
         model.train(xtr, y)
     else:
         model.train_nfold(xtr, y)
+
     # saving the model
-    model.save()
-    
+    model.save(output_directory)
+
     '''
     print('training second-level dataset subtype corpus...')
     xtr, y1, y2, _, list_classes, list_subclasses, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-multilevel.csv")
@@ -128,23 +129,23 @@ def train(embeddings_name, fold_count, architecture="gru", transformer=None, cas
         model.save()
     '''
 
-def train_and_eval(embeddings_name=None, fold_count=1, architecture="gru", transformer=None, cascaded=False): 
+def train_and_eval(embeddings_name=None, fold_count=1, architecture="gru", transformer=None, cascaded=False, output_directory=None):
     if cascaded:
-        return train_eval_cascaded(embeddings_name, fold_count, architecture=architecture, transformer=transformer)
+        return train_eval_cascaded(embeddings_name, fold_count, architecture=architecture, transformer=transformer, output_directory=output_directory)
 
     # classifier for deciding if we have a dataset or not in a sentence
-    train_and_eval_binary(embeddings_name, fold_count, architecture=architecture, transformer=transformer)
+    train_and_eval_binary(embeddings_name, fold_count, architecture=architecture, transformer=transformer, output_directory=output_directory)
 
     # classifier for deciding if the introduced dataset is a reuse of an existing one or is a new dataset
-    train_and_eval_reuse(embeddings_name, fold_count, architecture=architecture, transformer=transformer)
+    train_and_eval_reuse(embeddings_name, fold_count, architecture=architecture, transformer=transformer, output_directory=output_directory)
 
     # classifier for first level data type hierarchy
-    train_and_eval_primary(embeddings_name, fold_count, architecture=architecture, transformer=transformer)
+    train_and_eval_primary(embeddings_name, fold_count, architecture=architecture, transformer=transformer, output_directory=output_directory)
 
     # classifier for second level data type hierarchy (subtypes)
     #train_and_eval_secondary(embeddings_name, fold_count, architecture=architecture, transformer=transformer)
 
-def train_and_eval_binary(embeddings_name, fold_count, architecture="gru", transformer=None): 
+def train_and_eval_binary(embeddings_name, fold_count, architecture="gru", transformer=None, output_directory=None):
     print('loading dataset type corpus...')
     xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-binary.csv")
 
@@ -179,9 +180,9 @@ def train_and_eval_binary(embeddings_name, fold_count, architecture="gru", trans
     model.eval(x_test, y_test)
 
     # saving the model
-    model.save()
+    model.save(output_directory)
 
-def train_and_eval_reuse(embeddings_name, fold_count, architecture="gru", transformer=None): 
+def train_and_eval_reuse(embeddings_name, fold_count, architecture="gru", transformer=None, output_directory=None):
     print('loading dataset type corpus...')
     xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-reuse.csv")
 
@@ -216,9 +217,9 @@ def train_and_eval_reuse(embeddings_name, fold_count, architecture="gru", transf
     model.eval(x_test, y_test)
 
     # saving the model
-    model.save()
-    
-def train_and_eval_primary(embeddings_name, fold_count, architecture="gru", transformer=None): 
+    model.save(output_directory)
+
+def train_and_eval_primary(embeddings_name, fold_count, architecture="gru", transformer=None, output_directory=None):
     print('loading dataset type corpus...')
     xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv("data/textClassification/dataseer/all-multilevel.csv")
 
@@ -252,7 +253,7 @@ def train_and_eval_primary(embeddings_name, fold_count, architecture="gru", tran
     model.eval(x_test, y_test)
 
     # saving the model
-    model.save()
+    model.save(output_directory)
 
 def train_and_eval_secondary(embeddings_name, fold_count, architecture="gru", transformer=None): 
     print('training second-level dataset subtype corpus...')
@@ -506,6 +507,7 @@ def build_prior_class_distribution():
             "HuggingFace transformers hub will be used otherwise to fetch the model, see https://huggingface.co/models " + \
             "for model names"
     )
+    parser.add_argument("--output", help="Directory where to save a trained model.", default=DEFAULT_DATA_MODEL_PATH_TEXT_CLASSIFICATION)
 
     args = parser.parse_args()
 
@@ -515,25 +517,28 @@ def build_prior_class_distribution():
     embeddings_name = args.embedding
     cascaded = args.cascaded
     transformer = args.transformer
+    output = args.output
 
     architecture = args.architecture
     if architecture not in architectures:
         print('unknown model architecture, must be one of '+str(architectures))
 
-    if transformer == None and embeddings_name == None:
+    if transformer is None and embeddings_name is None:
         # default word embeddings
         embeddings_name = "glove-840B"
 
     if args.action == 'train':
         if args.fold_count < 1:
             raise ValueError("fold-count should be equal or more than 1")
 
-        train(embeddings_name=embeddings_name, fold_count=args.fold_count, architecture=architecture, transformer=transformer, cascaded=cascaded)
+        train(embeddings_name=embeddings_name, fold_count=args.fold_count, architecture=architecture,
+              transformer=transformer, cascaded=cascaded, output_directory=output)
 
     if args.action == 'train_eval':
         if args.fold_count < 1:
             raise ValueError("fold-count should be equal or more than 1")
-        y_test = train_and_eval(embeddings_name=embeddings_name, fold_count=args.fold_count, architecture=architecture, transformer=transformer, cascaded=cascaded)    
+        y_test = train_and_eval(embeddings_name=embeddings_name, fold_count=args.fold_count, architecture=architecture,
+                                transformer=transformer, cascaded=cascaded, output_directory=output)
 
     if args.action == 'classify':
         someTexts = ['Labeling yield and radiochemical purity was analyzed by instant thin layered chromatography (ITLC).', 

delft/applications/datasetTagger.py

@@ -8,7 +8,8 @@
 
 from delft.sequenceLabelling import Sequence
 from delft.sequenceLabelling.reader import load_data_and_labels_json_offsets
-from delft.utilities.misc import parse_number_ranges
+from delft.utilities.misc import DEFAULT_DATA_MODEL_PATH_SEQUENCE_LABELLING
+
 
 def configure(architecture, output_path=None, max_sequence_length=-1, batch_size=-1, embeddings_name=None, max_epoch=-1, use_ELMo=False):
     """
@@ -94,10 +95,7 @@ def train(embeddings_name=None, architecture='BidLSTM_CRF', transformer=None,
     print("training runtime: %s seconds " % runtime)
 
     # saving the model
-    if output_path:
-        model.save(output_path)
-    else:
-        model.save()
+    model.save(output_path)
 
 
 # split data, train a model and evaluate it
@@ -245,7 +243,7 @@ def annotate_text(texts, output_format, architecture='BidLSTM_CRF', features=Non
             "HuggingFace transformers hub will be used otherwise to fetch the model, see https://huggingface.co/models " + \
             "for model names"
     )
-    parser.add_argument("--output", help="Directory where to save a trained model.")
+    parser.add_argument("--output", help="Directory where to save a trained model.", default=DEFAULT_DATA_MODEL_PATH_SEQUENCE_LABELLING)
     parser.add_argument("--input", help="Grobid data file to be used for training (train action), for training and " +
                                         "evaluation (train_eval action) or just for evaluation (eval action).")
     parser.add_argument("--max-sequence-length", type=int, default=-1, help="max-sequence-length parameter to be used.")

delft/applications/grobidTagger.py

@@ -8,8 +8,7 @@
 
 from delft.sequenceLabelling import Sequence
 from delft.sequenceLabelling.reader import load_data_and_labels_crf_file
-from delft.sequenceLabelling.reader import load_data_crf_string
-from delft.utilities.misc import parse_number_ranges
+from delft.utilities.misc import DEFAULT_DATA_MODEL_PATH_SEQUENCE_LABELLING
 
 MODEL_LIST = ['affiliation-address', 'citation', 'date', 'header', 'name-citation', 'name-header', 'software', 'figure', 'table', 'reference-segmenter']
 
@@ -149,10 +148,7 @@ def train(model, embeddings_name=None, architecture=None, transformer=None, inpu
     print("training runtime: %s seconds " % (runtime))
 
     # saving the model
-    if output_path:
-        model.save(output_path)
-    else:
-        model.save()
+    model.save(output_path)
 
 
 # split data, train a GROBID model and evaluate it
@@ -209,10 +205,7 @@ def train_eval(model, embeddings_name=None, architecture='BidLSTM_CRF', transfor
     model.eval(x_eval, y_eval, features=f_eval)
 
     # saving the model (must be called after eval for multiple fold training)
-    if output_path:
-        model.save(output_path)
-    else:
-        model.save()
+    model.save(output_path)
 
 
 # split data, train a GROBID model and evaluate it
@@ -272,6 +265,7 @@ def annotate_text(texts, model, output_format, architecture='BidLSTM_CRF', featu
 
     return annotations
 
+
 class Tasks:
     TRAIN = 'train'
     TRAIN_EVAL = 'train_eval'
@@ -302,7 +296,8 @@ class Tasks:
     parser.add_argument("action", choices=actions)
     parser.add_argument("--fold-count", type=int, default=1, help="Number of fold to use when evaluating with n-fold "
                                                                   "cross validation.")
-    parser.add_argument("--architecture", help="Type of model architecture to be used, one of "+str(architectures))
+    parser.add_argument("--architecture", choices=architectures, help="Type of model architecture to be used",
+                        required=True)
     parser.add_argument("--use-ELMo", action="store_true", help="Use ELMo contextual embeddings") 
 
     # group_embeddings = parser.add_mutually_exclusive_group(required=False)
@@ -324,7 +319,7 @@ class Tasks:
             "HuggingFace transformers hub will be used otherwise to fetch the model, see https://huggingface.co/models " + \
             "for model names"
     )
-    parser.add_argument("--output", help="Directory where to save a trained model.")
+    parser.add_argument("--output", help="Directory where to save a trained model.", default=DEFAULT_DATA_MODEL_PATH_SEQUENCE_LABELLING)
     parser.add_argument("--input", help="Grobid data file to be used for training (train action), for training and " +
                                         "evaluation (train_eval action) or just for evaluation (eval action).")
     parser.add_argument("--max-sequence-length", type=int, default=-1, help="max-sequence-length parameter to be used.")