def train_and_eval(embeddings_name, fold_count, use_ELMo=False):
batch_size = 256
if use_ELMo:
batch_size = 20
model = Classifier('citations',
"gru",
list_classes=list_classes,
max_epoch=70,
fold_number=fold_count,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
batch_size=batch_size,
class_weights=class_weights)
print('loading citation sentiment corpus...')
xtr, y = load_citation_sentiment_corpus(
"data/textClassification/citations/citation_sentiment_corpus.txt")
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval_primary(embeddings_name,
fold_count,
use_ELMo=False,
use_BERT=False,
architecture="gru"):
print('loading dataset type corpus...')
xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv(
"data/textClassification/dataseer/all-1.csv")
# distinct values of classes
print(list_classes)
print(len(list_classes), "classes")
print(len(xtr), "texts")
print(len(y), "classes")
class_weights = None
batch_size = 256
maxlen = 300
if use_ELMo:
batch_size = 20
elif use_BERT:
batch_size = 50
# default bert model parameters
if architecture.find("bert") != -1:
batch_size = 32
maxlen = 100
model = Classifier('dataseer',
model_type=architecture,
list_classes=list_classes,
max_epoch=100,
fold_number=fold_count,
patience=10,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
use_BERT=use_BERT,
batch_size=batch_size,
maxlen=maxlen,
class_weights=class_weights)
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
print(len(x_train), "train texts")
print(len(y_train), "train classes")
print(len(x_test), "eval texts")
print(len(y_test), "eval classes")
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval(embeddings_name,
fold_count,
use_ELMo=False,
use_BERT=False,
architecture="gru"):
print('loading binary software use dataset...')
xtr, y = load_software_use_corpus_json(
"data/textClassification/software/software-use.json.gz")
nb_used = 0
for the_class in y:
if the_class[1] == 1.0:
nb_used += 1
nb_unused = len(y) - nb_used
print("\ttotal:", len(y))
print("\tused:", nb_used)
print("\tnot used:", nb_unused)
model_name = 'software_use'
class_weights = None
if use_ELMo:
model_name += '-with_ELMo'
elif use_BERT:
model_name += '-with_BERT'
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
batch_size, maxlen = configure(architecture, use_BERT, use_ELMo)
print(list_classes)
model = Classifier(model_name,
model_type=architecture,
list_classes=list_classes,
max_epoch=100,
fold_number=fold_count,
patience=10,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
use_BERT=use_BERT,
batch_size=batch_size,
maxlen=maxlen,
class_weights=class_weights)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval_binary(embeddings_name,
fold_count,
architecture="gru",
transformer=None):
print('loading multiclass software context dataset...')
xtr, y = load_software_context_corpus_json(
"data/textClassification/software/software-contexts.json.gz")
report_training_contexts(y)
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
for class_rank in range(len(list_classes)):
model_name = 'software_context_' + list_classes[
class_rank] + '_' + architecture
class_weights = None
batch_size, maxlen, patience, early_stop, max_epoch = configure(
architecture)
y_train_class_rank = [[1, 0] if y[class_rank] == 1.0 else [0, 1]
for y in y_train]
y_test_class_rank = [[1, 0] if y[class_rank] == 1.0 else [0, 1]
for y in y_test]
y_train_class_rank = np.array(y_train_class_rank)
y_test_class_rank = np.array(y_test_class_rank)
list_classes_rank = [
list_classes[class_rank], "not_" + list_classes[class_rank]
]
model = Classifier(model_name,
architecture=architecture,
list_classes=list_classes_rank,
max_epoch=max_epoch,
fold_number=fold_count,
patience=patience,
use_roc_auc=True,
embeddings_name=embeddings_name,
batch_size=batch_size,
maxlen=maxlen,
early_stop=early_stop,
class_weights=class_weights,
transformer_name=transformer)
if fold_count == 1:
model.train(x_train, y_train_class_rank)
else:
model.train_nfold(x_train, y_train_class_rank)
model.eval(x_test, y_test_class_rank)
def train_and_eval_reuse(embeddings_name,
fold_count,
architecture="gru",
transformer=None):
print('loading dataset type corpus...')
xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv(
"data/textClassification/dataseer/all-reuse.csv")
# distinct values of classes
print(list_classes)
print(len(list_classes), "classes")
print(len(xtr), "texts")
print(len(y), "classes")
batch_size, maxlen, patience, early_stop, max_epoch = configure(
architecture)
class_weights = {0: 1.5, 1: 1.}
model = Classifier('dataseer-reuse_' + architecture,
architecture=architecture,
list_classes=list_classes,
max_epoch=max_epoch,
fold_number=fold_count,
use_roc_auc=True,
embeddings_name=embeddings_name,
batch_size=batch_size,
maxlen=maxlen,
patience=patience,
early_stop=early_stop,
class_weights=class_weights,
transformer_name=transformer)
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
print(len(x_train), "train texts")
print(len(y_train), "train classes")
print(len(x_test), "eval texts")
print(len(y_test), "eval classes")
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval(embeddings_name,
fold_count,
architecture="gru",
transformer=None):
print('loading binary software use dataset...')
xtr, y = load_software_use_corpus_json(
"data/textClassification/software/software-use.json.gz")
nb_used = 0
for the_class in y:
if the_class[1] == 1.0:
nb_used += 1
nb_unused = len(y) - nb_used
print("\ttotal:", len(y))
print("\tused:", nb_used)
print("\tnot used:", nb_unused)
model_name = 'software_use_' + architecture
class_weights = None
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
batch_size, maxlen, patience, early_stop, max_epoch = configure(
architecture)
print(list_classes)
model = Classifier(model_name,
architecture=architecture,
list_classes=list_classes,
max_epoch=max_epoch,
fold_number=fold_count,
patience=patience,
use_roc_auc=True,
embeddings_name=embeddings_name,
batch_size=batch_size,
maxlen=maxlen,
early_stop=early_stop,
class_weights=class_weights,
transformer_name=transformer)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval(embeddings_name,
fold_count,
architecture="gru",
transformer=None):
print('loading multiclass software context dataset...')
xtr, y = load_software_context_corpus_json(
"data/textClassification/software/software-contexts.json.gz")
report_training_contexts(y)
model_name = 'software_context_' + architecture
class_weights = None
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
batch_size, maxlen, patience, early_stop, max_epoch = configure(
architecture)
print(list_classes)
model = Classifier(model_name,
architecture=architecture,
list_classes=list_classes,
max_epoch=max_epoch,
fold_number=fold_count,
patience=patience,
use_roc_auc=True,
embeddings_name=embeddings_name,
batch_size=batch_size,
maxlen=maxlen,
early_stop=early_stop,
class_weights=class_weights,
transformer_name=transformer)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval(embeddings_name,
fold_count,
use_ELMo=False,
use_BERT=False,
architecture="gru"):
batch_size, maxlen = configure(architecture, use_BERT, use_ELMo)
maxlen = 150
model = Classifier('citations',
model_type=architecture,
list_classes=list_classes,
max_epoch=100,
fold_number=fold_count,
patience=10,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
use_BERT=use_BERT,
batch_size=batch_size,
maxlen=maxlen,
class_weights=class_weights)
print('loading citation sentiment corpus...')
xtr, y = load_citation_sentiment_corpus(
"data/textClassification/citations/citation_sentiment_corpus.txt")
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
def train_and_eval(embeddings_name,
fold_count,
architecture="gru",
transformer=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,
use_roc_auc=True,
embeddings_name=embeddings_name,
batch_size=batch_size,
maxlen=maxlen,
patience=patience,
early_stop=early_stop,
class_weights=class_weights,
transformer_name=transformer)
print('loading citation sentiment corpus...')
xtr, y = load_citation_sentiment_corpus(
"data/textClassification/citations/citation_sentiment_corpus.txt")
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
# saving the model
model.save()
model.eval(x_test, y_test)
def train_eval_cascaded(embeddings_name,
fold_count,
use_ELMo=False,
use_BERT=False,
architecture="gru"):
# general setting of parameters
class_weights = None
batch_size = 256
maxlen = 300
if use_ELMo:
batch_size = 20
elif use_BERT:
batch_size = 50
# default bert model parameters
if architecture.find("bert") != -1:
batch_size = 32
maxlen = 100
# first binary classifier: dataset or no_dataset
xtr, y, _, _, list_classes, _, _ = load_dataseer_corpus_csv(
"data/textClassification/dataseer/all-binary.csv")
print(list_classes)
model_binary = Classifier('dataseer-binary',
model_type=architecture,
list_classes=list_classes,
max_epoch=100,
fold_number=fold_count,
patience=10,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
use_BERT=use_BERT,
batch_size=batch_size,
maxlen=maxlen,
class_weights=class_weights)
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(xtr, y, 0.9)
if fold_count == 1:
model_binary.train(x_train, y_train)
else:
model_binary.train_nfold(x_train, y_train)
model_binary.eval(x_test, y_test)
x_test_binary = x_test
y_test_binary = y_test
# second, the first level datatype taxonomy for sentences classified as dataset
xtr, y_classes, y_subclasses, y_leafclasses, list_classes, list_subclasses, list_leaf_classes = load_dataseer_corpus_csv(
"data/textClassification/dataseer/all-1.csv")
# ignore the no_dataset, ignore the first eval set, build first level classifier
ind = list_classes.index('no_dataset')
to_remove = vectorizer(ind, len(list_classes))
x_train, y_train = filter_exclude_class(xtr, y_classes, to_remove)
y_train2 = np.zeros(shape=(len(y_train), len(list_classes) - 1))
for i in range(0, len(y_train)):
y_train2[i] = np.delete(y_train[i], ind)
y_train = y_train2
list_classes.remove('no_dataset')
model_first = Classifier('dataseer-first',
model_type=architecture,
list_classes=list_classes,
max_epoch=100,
fold_number=fold_count,
patience=10,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
use_BERT=use_BERT,
batch_size=batch_size,
maxlen=maxlen,
class_weights=class_weights)
if fold_count == 1:
model_first.train(x_train, y_train)
else:
model_first.train_nfold(x_train, y_train)
model_first.eval(x_test, y_test)
# eval by cascading
result_binary = model_binary.predict(x_test_binary,
output_format='default')
result_first = model_first.predict(x_test, output_format='default')
# select sequences classified as dataset
result_intermediate = np.asarray(
[np.argmax(line) for line in result_binary])
def vectorize(index, size):
result = np.zeros(size)
if index < size:
result[index] = 1
return result
result_binary = np.array(
[vectorize(xi, len(list_classes)) for xi in result_intermediate])
def train_and_eval_secondary(embeddings_name,
fold_count,
use_ELMo=False,
use_BERT=False,
architecture="gru"):
print('training second-level dataset subtype corpus...')
xtr, y1, y2, _, list_classes, list_subclasses, _ = load_dataseer_corpus_csv(
"data/textClassification/dataseer/all-1.csv")
# aggregate by class, we will have one training set per class
print(list_classes)
print(list_subclasses)
print(len(list_classes), "classes")
print(len(list_subclasses), "sub-classes")
class_weights = None
batch_size = 256
maxlen = 300
if use_ELMo:
batch_size = 20
elif use_BERT:
batch_size = 50
# default bert model parameters
if architecture.find("bert") != -1:
batch_size = 32
maxlen = 100
datatypes_y = {}
datatypes_xtr = {}
datatypes_list_subclasses = {}
for i in range(0, len(xtr)):
#print(np.where(y2[i] == 1))
ind1 = np.where(y1[i] == 1)[0][0]
ind2 = np.where(y2[i] == 1)[0][0]
#print(ind2)
datatype = list_classes[ind1]
datasubtype = list_subclasses[ind2]
#print(str(xtr[i]), datatype, datasubtype)
if datatype in datatypes_y:
datatypes_y[datatype].append(datasubtype)
datatypes_xtr[datatype].append(xtr[i])
if not datasubtype in datatypes_list_subclasses[datatype]:
datatypes_list_subclasses[datatype].append(datasubtype)
else:
datatypes_y[datatype] = []
datatypes_y[datatype].append(datasubtype)
datatypes_xtr[datatype] = []
datatypes_xtr[datatype].append(xtr[i])
datatypes_list_subclasses[datatype] = []
datatypes_list_subclasses[datatype].append(datasubtype)
print(datatypes_list_subclasses)
for the_class in list_classes:
print('\ntraining', the_class)
if not the_class in datatypes_list_subclasses:
print('no subclass for', the_class)
continue
if len(datatypes_list_subclasses[the_class]) <= 1:
print('only one subclass for', the_class)
continue
if len(datatypes_list_subclasses[the_class]
) == 2 and 'nan' in datatypes_list_subclasses[the_class]:
continue
if the_class == 'Protein Data':
continue
print('subtypes to be classified:',
datatypes_list_subclasses[the_class])
model_name = 'dataseer-' + the_class
if use_ELMo:
model_name += '-with_ELMo'
elif use_BERT:
model_name += '-with_BERT'
model = Classifier(model_name,
model_type=architecture,
list_classes=datatypes_list_subclasses[the_class],
max_epoch=100,
fold_number=fold_count,
patience=10,
use_roc_auc=True,
embeddings_name=embeddings_name,
use_ELMo=use_ELMo,
use_BERT=use_BERT,
batch_size=batch_size,
maxlen=maxlen,
class_weights=class_weights)
# we need to vectorize the y according to the actual list of classes
local_y = []
for the_y in datatypes_y[the_class]:
the_ind = datatypes_list_subclasses[the_class].index(the_y)
local_y.append(
vectorizer(the_ind, len(datatypes_list_subclasses[the_class])))
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(
np.asarray(datatypes_xtr[the_class]), np.asarray(local_y), 0.9)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
def train_and_eval_secondary(embeddings_name,
fold_count,
architecture="gru",
transformer=None):
print('training second-level dataset subtype corpus...')
xtr, y1, y2, _, list_classes, list_subclasses, _ = load_dataseer_corpus_csv(
"data/textClassification/dataseer/all-multilevel.csv")
# aggregate by class, we will have one training set per class
print(list_classes)
print(list_subclasses)
print(len(list_classes), "classes")
print(len(list_subclasses), "sub-classes")
class_weights = None
batch_size, maxlen, patience, early_stop, max_epoch = configure(
architecture)
datatypes_y = {}
datatypes_xtr = {}
datatypes_list_subclasses = {}
for i in range(0, len(xtr)):
#print(np.where(y2[i] == 1))
ind1 = np.where(y1[i] == 1)[0][0]
ind2 = np.where(y2[i] == 1)[0][0]
#print(ind2)
datatype = list_classes[ind1]
datasubtype = list_subclasses[ind2]
#print(str(xtr[i]), datatype, datasubtype)
if datatype in datatypes_y:
datatypes_y[datatype].append(datasubtype)
datatypes_xtr[datatype].append(xtr[i])
if not datasubtype in datatypes_list_subclasses[datatype]:
datatypes_list_subclasses[datatype].append(datasubtype)
else:
datatypes_y[datatype] = []
datatypes_y[datatype].append(datasubtype)
datatypes_xtr[datatype] = []
datatypes_xtr[datatype].append(xtr[i])
datatypes_list_subclasses[datatype] = []
datatypes_list_subclasses[datatype].append(datasubtype)
print(datatypes_list_subclasses)
for the_class in list_classes:
print('\ntraining', the_class)
if not the_class in datatypes_list_subclasses:
print('no subclass for', the_class)
continue
if len(datatypes_list_subclasses[the_class]) <= 1:
print('only one subclass for', the_class)
continue
if len(datatypes_list_subclasses[the_class]
) == 2 and 'nan' in datatypes_list_subclasses[the_class]:
continue
if the_class == 'Protein Data':
continue
print('subtypes to be classified:',
datatypes_list_subclasses[the_class])
model_name = 'dataseer-' + the_class + "_" + architecture
model = Classifier(model_name,
architecture=architecture,
list_classes=datatypes_list_subclasses[the_class],
max_epoch=max_epoch,
fold_number=fold_count,
use_roc_auc=True,
embeddings_name=embeddings_name,
batch_size=batch_size,
maxlen=maxlen,
patience=patience,
early_stop=early_stop,
class_weights=class_weights,
transformer_name=transformer)
# we need to vectorize the y according to the actual list of classes
local_y = []
for the_y in datatypes_y[the_class]:
the_ind = datatypes_list_subclasses[the_class].index(the_y)
local_y.append(
vectorizer(the_ind, len(datatypes_list_subclasses[the_class])))
# segment train and eval sets
x_train, y_train, x_test, y_test = split_data_and_labels(
np.asarray(datatypes_xtr[the_class]), np.asarray(local_y), 0.9)
if fold_count == 1:
model.train(x_train, y_train)
else:
model.train_nfold(x_train, y_train)
model.eval(x_test, y_test)
# saving the model
model.save()
