def finetune(input_path, output_model_dir, text_col, label_col, model_dir, gpu, with_header):
'''Train a new model or finetune an existing model with labels, output fine-tuned model'''
# assign GPU
os.environ['CUDA_VISIBLE_DEVICES'] = gpu
dataset = csv_handler.csv_readlines(input_path)
header = None
if with_header == True:
header = dataset[0]
dataset = dataset[1:]
print("Loading source model from %s ...\n" % (model_dir))
model = BertModel(model_dir)
text_col = text_col - 1
label_col = label_col - 1
labels = transform.map_func(range(len(dataset)), lambda i : [i, dataset[i][text_col], dataset[i][label_col]])
print("Fine-tuning with input labels")
model.train(labels)
model.checkpoint(output_model_dir)
print("Finished. Fine-tuned model is ready at " + output_model_dir)
def transform(input_path, output_path):
dataset = csv_handler.csv_readlines(input_path)
dataset = dataset[1:]
dataset = transformer.map_func(
range(len(dataset)), lambda i: (i, dataset[i][0], 1
if dataset[i][1] == "True" else 0))
csv_handler.csv_writelines(output_path, dataset)
def extractor(anno_dir, id_to_file, paper_id):
file_path = anno_dir + id_to_file[paper_id]
label_sent_dataset = csv_handler.csv_readlines(file_path, delimit='\t')
indexed_result = transformer.indexleft_func(label_sent_dataset)
final = transformer.map_func(
indexed_result, lambda p:
(paper_id + "_" + str(p[0]), p[1][1], p[1][0]))
return final
def bert_estimate(input_path, text_col, output_path, model_dir, gpu,
with_header):
# assign GPU
os.environ['CUDA_VISIBLE_DEVICES'] = gpu
# load model
if model_dir == '':
model_dir = os.path.join(os.environ['FACTMINE_HOME'],
'cache/company_model')
print("Loading model from %s ...\n" % (model_dir))
model = BertModel(model_dir)
print("Calculating negative score, positive score, and argmax class...\n")
dataset = csv_handler.csv_readlines(input_path)
header = None
if with_header == True:
header = dataset[0]
dataset = dataset[1:]
text_col = text_col - 1
texts = transform.map_func(dataset, lambda row: row[text_col])
preds = model.predict(texts)
assert (len(dataset) == len(preds))
# apply sigmoid
preds = transform.map_func(
preds, lambda quad:
[util.sigmoid(quad[0]),
util.sigmoid(quad[1]), quad[2], quad[3]])
for i in range(len(dataset)):
dataset[i].append(preds[i][0])
dataset[i].append(preds[i][1])
dataset[i].append(preds[i][2])
if with_header == True:
header.append('score_0')
header.append('score_1')
header.append('arg_class')
dataset = [header] + dataset
csv_handler.csv_writelines(output_path, dataset)
if output_path != "":
print("Finished, Results are ready at %s " % (output_path))
def extract(data, typename):
dataset = csv_handler.csv_readlines("./dataset/" + data + "set.csv")
dataset = transformer.indexleft_func(dataset)
dataset = transformer.map_func(dataset, lambda row:
(row[0], row[1][1], row[1][2]))
output_path = "./" + data + ".csv"
def e_func(triplet):
label = 0
if triplet[2] == typename:
label = 1
return (triplet[0], triplet[1], label)
final = transformer.map_func(dataset, lambda triplet: e_func(triplet))
csv_handler.csv_writelines(output_path, final)
def max_balancer(input_csv_path, output_csv_path='./output.csv'):
dataset = csv_handler.csv_readlines(input_csv_path)
pos_dataset = transform.filter_func(dataset, lambda row: row[2] == '1')
neg_dataset = transform.filter_func(dataset, lambda row: row[2] == '0')
assert (len(pos_dataset) <= len(neg_dataset))
sampler = Sampler()
neg_dataset = sampler.sample_rows(neg_dataset, len(pos_dataset))
pos_ids = transform.map_func(pos_dataset, lambda row: row[0])
neg_ids = transform.map_func(neg_dataset, lambda row: row[0])
select_id_set = set(pos_ids + neg_ids)
final = transform.filter_func(dataset, lambda row: row[0] in select_id_set)
csv_handler.csv_writelines(output_csv_path, final)
def extract(data, typename):
dataset = csv_handler.csv_readlines("./dataset/" + data + "_raw.csv")
output_path = "./" + data + ".csv"
def func_1(triplet):
label = 0
if triplet[2] == typename:
label = 1
return (triplet[0], triplet[1], label)
e_func = func_1
def func_2(triplet):
label = 1
if triplet[2] == typename:
label = 0
return (triplet[0], triplet[1], label)
if typename == 'NoArgument':
e_func = func_2
final = transformer.map_func(dataset, lambda triplet: e_func(triplet))
csv_handler.csv_writelines(output_path, final)
def evaluate(input_path, col_true, col_pred, metric, output_path, with_header):
'''evaluate the quality of predictions with a metric (f1 by default), and output the metric scores'''
result = []
dataset = csv_handler.csv_readlines(input_path)
if with_header == True:
dataset = dataset[1:]
col_true = int(col_true) - 1
col_pred = int(col_pred) - 1
y_true = transform.map_func(dataset, lambda row : int(row[col_true]))
y_pred = transform.map_func(dataset, lambda row : int(row[col_pred]))
def check_validity(class_array):
for cls in class_array:
assert(cls == 0 or cls == 1)
check_validity(y_true)
check_validity(y_pred)
support_set = {'f1', 'accuracy', 'cohen', 'quad'}
if metric not in support_set:
sys.exit('please specify a valid metric in terms of f1, accuracy, cohen, or quad (i.e. precision_recall_fscore_support)')
elif metric == 'f1':
result.append(['f1'])
result.append([f1_score(y_true, y_pred)])
elif metric == 'accuracy':
result.append(['accuracy'])
result.append([accuracy_score(y_true, y_pred)])
elif metric == 'cohen':
result.append([cohen_kappa_score(y_true, y_pred)])
elif metric == 'quad':
(precision, recall, fscore, support) = precision_recall_fscore_support(y_true, y_pred)
result.append(['class', 'precision', 'recall', 'fscore', 'support'])
result.append([0, precision[0], recall[0], fscore[0], support[0]])
result.append([1, precision[1], recall[1], fscore[1], support[1]])
csv_handler.csv_writelines(output_path, result)
def write_preds(ori_preds, labels, dev_path , output_path = "./tmp.csv"):
eval_examples = csv_handler.csv_readlines(dev_path)
assert(len(ori_preds) == len(labels))
assert(len(ori_preds) == len(eval_examples))
# append header
header = ['id', 'sent', 'label']
for i in range(len(ori_preds[0])):
header.append('p' + str(i))
header.append('pred')
final = []
final.append(header)
# append data
for i in range(len(labels)):
row = []
sent_id = eval_examples[i][0]
sent = eval_examples[i][1]
sent_label = int(eval_examples[i][2])
assert(sent_label == labels[i])
row.append(sent_id)
row.append(sent)
row.append(sent_label)
for j in range(len(ori_preds[i])):
row.append(ori_preds[i][j])
pred_label = np.argmax(ori_preds[i])
row.append(pred_label)
final.append(row)
csv_handler.csv_writelines(output_path, final)
from sklearn.feature_extraction.text import TfidfTransformer
import os
file_path = os.path.dirname(os.path.abspath(__file__))
import sys
sys.path.insert(0, '../../pyfunctor')
import csv_handler as csv_handler
import transform as transformer
from sklearn import metrics
import time
import gensim
from pytorch_pretrained_bert import BertTokenizer
import os
start_time = time.time()
train_set = csv_handler.csv_readlines(sys.argv[1])
dev_set = csv_handler.csv_readlines(sys.argv[2])
output_path = sys.argv[3]
if os.path.exists(output_path):
os.remove(output_path)
seed = int(sys.argv[4])
def sep(dataset):
sents = transformer.map_func(dataset, lambda triplet: triplet[1])
labels = transformer.map_func(dataset, lambda triplet: (int)(triplet[2]))
return (sents, labels)
sys.path.insert(0, "../../pyfunctor")
import csv_handler as csv_handler
import transform as transformer
import csv
class CSV_Split(csv_handler.CSV_Handler):
def __init__(self, dataset, seed=0):
self.seed = seed
self.dataset = dataset
input_dir = "./DEEPTip/dataset/"
# paragraph
sent_n = csv_handler.csv_readlines(input_dir + "sent_tip.neg",
delimit='\t',
quoter=csv.QUOTE_NONE)
sent_n = transformer.map_func(sent_n, lambda p: (p[0], p[1], '0'))
sent_p = csv_handler.csv_readlines(input_dir + "sent_tip.pos",
delimit='\t',
quoter=csv.QUOTE_NONE)
sent_p = transformer.map_func(sent_p, lambda p: (p[0], p[1], '1'))
sent = sent_n + sent_p
for row in sent:
assert (len(row) == 3)
ids = set(transformer.map_func(sent, lambda p: p[0]))
assert (len(ids) == len(sent))
splitter = CSV_Split(sent)
root.tag
sents = []
for child in root:
tid = child.attrib['id']
sentence = ""
num_words = len(child)
for i in range(num_words):
sentence += child[i].text
if i < num_words - 2:
sentence += " "
sents.append((tid, sentence))
import csv_handler as csv_handler
golds = csv_handler.csv_readlines(gold_path, delimit='\t')
import transform as transformer
assert(len(sents) == len(golds))
for i in range(len(sents)):
assert(sents[i][0] == golds[i][0])
final = transformer.map_func(range(len(sents)), lambda i : (sents[i][0], sents[i][1], golds[i][1]))
import csv_handler as csv_handler
class CSV_Split(csv_handler.CSV_Handler):
def __init__(self, dataset, seed = 0):
self.seed = seed
self.dataset = dataset
splitter = CSV_Split(final)
splitter.csv_split(0.2, "dev.csv", "train.csv")
import transform as transform
import csv_handler as csv_handler
if __name__ == "__main__":
result = []
first_file = sys.argv[1]
first_idx = sys.argv[2].split(',')
first_idx = transform.map_func(first_idx, lambda idx: int(idx))
second_file = sys.argv[3]
second_idx = sys.argv[4].split(',')
second_idx = transform.map_func(second_idx, lambda idx: int(idx))
output_csv_file = sys.argv[5]
first_dataset = csv_handler.csv_readlines(first_file)
result = transform.map_func(first_dataset,
lambda row: [row[idx] for idx in first_idx])
second_dataset = csv_handler.csv_readlines(second_file)
second_result = transform.map_func(
second_dataset, lambda row: [row[idx] for idx in second_idx])
assert (len(first_dataset) == len(second_dataset))
final = transform.map_func(zip(result, second_result),
lambda p: p[0] + p[1])
csv_handler.csv_writelines(output_csv_file, final)
import sys
root_directory = '../../'
sys.path.insert(0, root_directory + "pyfunctor")
import transform as transformer
import csv_handler as csv_handler
import json_handler as json_handler
review_input_path = "../FUNNY/yelp_academic_dataset_review.csv"
review_output_path = "./all.csv"
dataset = csv_handler.csv_readlines(review_input_path)
# get idx for review_id, text, and funny_count
def get_triplet_idx(header):
idx_id = header.index('review_id')
idx_text = header.index('text')
idx_count = header.index('funny')
return (idx_id, idx_text, idx_count)
def selector(row, idx_id, idx_text, idx_count):
r_id = row[idx_id]
text = row[idx_text]
funny_count = int(row[idx_count])
is_funny = None
if (funny_count >= 5):
is_funny = 1
elif (funny_count == 0):
import sys sys.path.insert(0, '../../pyfunctor') import transform as transformer import csv_handler as csv_handler input_file=sys.argv[1] output_file=sys.argv[2] dataset = csv_handler.csv_readlines(input_file) dataset = transformer.indexleft_func(dataset) final = transformer.map_func(dataset, lambda p: (p[0], p[1][2], int(p[1][0]) - 1)) csv_handler.csv_writelines(output_file, final)
def distill(input_path):
dataset = csv_handler.csv_readlines(input_path, delimit='\t')
return dataset
def __init__(self, csv_file_path):
dataset = csv_handler.csv_readlines(csv_file_path)
labels = transform.map_func(dataset, lambda t: t[2])
self.weight_map = weight_class(labels)
import csv_handler as csv_handler
import transform as transformer
data_root = "./dataset"
data_dir = data_root + "/data/complete/"
topics = [
'abortion', 'cloning', 'death_penalty', 'gun_control',
'marijuana_legalization', 'minimum_wage', 'nuclear_energy',
'school_uniforms'
]
dataset = []
for tp in topics:
file_path = data_dir + tp + ".tsv"
records = csv_handler.csv_readlines(file_path,
delimit='\t',
quoter=csv.QUOTE_NONE)
records = records[1:]
def row_functor(i, records):
assert (i < len(records))
row = records[i]
rid = row[0] + "_" + str(i)
sent = row[4]
label = row[5]
split = row[6]
return (rid, sent, label, split)
records = transformer.map_func(range(len(records)),
csv_input_path = sys.argv[1]
y_true_col = int(sys.argv[2])
y_pred_col = int(sys.argv[3])
num_threds = int(sys.argv[4])
csv_output_path = sys.argv[5]
print_header = '1'
if len(sys.argv) > 6:
print_header = sys.argv[6]
assert (print_header == '1' or print_header == '0')
thred_method = "min_max_even"
if os.path.exists(csv_output_path):
os.remove(csv_output_path)
csv_dataset = csv_handler.csv_readlines(csv_input_path)
y_true = transform.map_func(csv_dataset, lambda row: int(row[y_true_col]))
y_pred_score = transform.map_func(csv_dataset,
lambda row: float(row[y_pred_col]))
#y_pred_score = transform.map_func(y_pred_score, lambda score : 1 / (1 + math.exp(-score)))
thred_col = []
if thred_method == "min_max_even":
thred_col = get_threds_by_min_max_even(y_pred_score, num_threds)
else: # sorted score even slot
thred_col = get_threds_by_sorted_score_equal_length(
y_pred_score, num_threds)
examples = []
for (i, txt) in enumerate(batch):
guid = "%s" % (i)
text_a = txt
label = self.dummy_label
examples.append(
InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
return examples
if __name__ == "__main__":
data_path = sys.argv[1]
model_dir = sys.argv[2]
output_path = sys.argv[3]
# load test dataset
raw_dataset = csv_handler.csv_readlines(data_path)
ids = transform.map_func(raw_dataset, lambda row : row[0])
texts = transform.map_func(raw_dataset, lambda row : row[1])
# load model
model = BertModel(model_dir)
pred = model.predict(texts, 100)
assert(len(ids) == len(pred))
output = transform.map_func(range(len(ids)), lambda idx : [ids[idx]] + pred[idx])
csv_handler.csv_writelines(output_path, output)
def distill(input_path, label_map):
dataset = csv_handler.csv_readlines(input_path, delimit='\t')
dataset = transform.filter_func(dataset, lambda row: len(row) >= 5)
token_label = transform.map_func(
dataset, lambda row: [row[0].strip(), label_map[row[4].strip()[0]]])
return token_label
