train.py

# Will be based on
# ConstructiveLoss function.
#
# https://github.com/UKPLab/sentence-transformers/blob/master/examples/training/quora_duplicate_questions/training_OnlineContrastiveLoss.py

from torch.utils.data import DataLoader
from sentence_transformers import losses, util
from sentence_transformers import LoggingHandler, SentenceTransformer, evaluation
from sentence_transformers.readers import InputExample
import logging
from datetime import datetime
import csv
import os
from zipfile import ZipFile
import random

#### Just some code to print debug information to stdout
logging.basicConfig(format='%(asctime)s - %(message)s',
                    datefmt='%Y-%m-%d %H:%M:%S',
                    level=logging.INFO,
                    handlers=[LoggingHandler()])
logger = logging.getLogger(__name__)
#### /print debug information to stdout


#As base model, we use DistilBERT-base that was pre-trained on NLI and STSb data
model = SentenceTransformer('sentence-transformers/paraphrase-albert-base-v2')
num_epochs = 12
# Smaller is generally better more accurate results.
train_batch_size = 5

#As distance metric, we use cosine distance (cosine_distance = 1-cosine_similarity)
distance_metric = losses.SiameseDistanceMetric.COSINE_DISTANCE

#Negative pairs should have a distance of at least 0.5
margin = 0.5

dataset_path = "data_set_training.csv"
model_save_path = 'output/training_OnlineConstrativeLoss-'+datetime.now().strftime("%Y-%m-%d_%H-%M-%S")

os.makedirs(model_save_path, exist_ok=True)

######### Read train data  ##########
# Read train data
train_samples = []
with open(dataset_path, encoding='utf8') as fIn:
    reader = csv.DictReader(fIn, delimiter='|', quoting=csv.QUOTE_NONE)
    for row in reader:
        sample = InputExample(texts=[row['ADDRESS1'], row['ADDRESS2']], label=int(row['ARE_SAME']))
        train_samples.append(sample)


train_dataloader = DataLoader(train_samples, shuffle=True, batch_size=train_batch_size)
train_loss = losses.OnlineContrastiveLoss(model=model, distance_metric=distance_metric, margin=margin)


################### Development  Evaluators ##################
# We add 3 evaluators, that evaluate the model on Duplicate Questions pair classification,
# Duplicate Questions Mining, and Duplicate Questions Information Retrieval
#evaluators = []

###### Classification ######
# Given (quesiton1, question2), is this a duplicate or not?
# The evaluator will compute the embeddings for both questions and then compute
# a cosine similarity. If the similarity is above a threshold, we have a duplicate.
dev_sentences1 = []
dev_sentences2 = []
dev_labels = []
with open( "dev_set_training.csv", encoding='utf8') as fIn:
    reader = csv.DictReader(fIn, delimiter='|', quoting=csv.QUOTE_NONE)
    for row in reader:
        dev_sentences1.append(row['ADDRESS1'])
        dev_sentences2.append(row['ADDRESS2'])
        dev_labels.append(int(row['ARE_SAME']))


binary_acc_evaluator = evaluation.BinaryClassificationEvaluator(dev_sentences1, dev_sentences2, dev_labels)
#evaluators.append(binary_acc_evaluator)



###### Duplicate Questions Mining ######
# Given a large corpus of questions, identify all duplicates in that corpus.

# For faster processing, we limit the development corpus to only 10,000 sentences.
#max_dev_samples = 10000
#dev_sentences = {}
#dev_duplicates = []
#with open("dev_corpus.csv", encoding='utf8') as fIn:
#    reader = csv.DictReader(fIn, delimiter='|', quoting=csv.QUOTE_NONE)
#    for row in reader:
#        dev_sentences[row['qid']] = row['question']
#
#        if len(dev_sentences) >= max_dev_samples:
#            break
#
#with open(os.path.join(dataset_path, "duplicate-mining/dev_duplicates.tsv"), encoding='utf8') as fIn:
#    reader = csv.DictReader(fIn, delimiter='\t', quoting=csv.QUOTE_NONE)
#    for row in reader:
#        if row['qid1'] in dev_sentences and row['qid2'] in dev_sentences:
#            dev_duplicates.append([row['qid1'], row['qid2']])
#
#
## The ParaphraseMiningEvaluator computes the cosine similarity between all sentences and
## extracts a list with the pairs that have the highest similarity. Given the duplicate
## information in dev_duplicates, it then computes and F1 score how well our duplicate mining worked
#paraphrase_mining_evaluator = evaluation.ParaphraseMiningEvaluator(dev_sentences, dev_duplicates, name='dev')
#evaluators.append(paraphrase_mining_evaluator)
#
#
####### Duplicate Questions Information Retrieval ######
## Given a question and a large corpus of thousands questions, find the most relevant (i.e. duplicate) question
## in that corpus.
#
## For faster processing, we limit the development corpus to only 10,000 sentences.
#max_corpus_size = 100000
#
#ir_queries = {}             #Our queries (qid => question)
#ir_needed_qids = set()      #QIDs we need in the corpus
#ir_corpus = {}              #Our corpus (qid => question)
#ir_relevant_docs = {}       #Mapping of relevant documents for a given query (qid => set([relevant_question_ids])
#
#with open(os.path.join(dataset_path, 'information-retrieval/dev-queries.tsv'), encoding='utf8') as fIn:
#    next(fIn) #Skip header
#    for line in fIn:
#        qid, query, duplicate_ids = line.strip().split('\t')
#        duplicate_ids = duplicate_ids.split(',')
#        ir_queries[qid] = query
#        ir_relevant_docs[qid] = set(duplicate_ids)
#
#        for qid in duplicate_ids:
#            ir_needed_qids.add(qid)
#
## First get all needed relevant documents (i.e., we must ensure, that the relevant questions are actually in the corpus
#distraction_questions = {}
#with open(os.path.join(dataset_path, 'information-retrieval/corpus.tsv'), encoding='utf8') as fIn:
#    next(fIn) #Skip header
#    for line in fIn:
#        qid, question = line.strip().split('\t')
#
#        if qid in ir_needed_qids:
#            ir_corpus[qid] = question
#        else:
#            distraction_questions[qid] = question
#
## Now, also add some irrelevant questions to fill our corpus
#other_qid_list = list(distraction_questions.keys())
#random.shuffle(other_qid_list)
#
#for qid in other_qid_list[0:max(0, max_corpus_size-len(ir_corpus))]:
#    ir_corpus[qid] = distraction_questions[qid]
#
##Given queries, a corpus and a mapping with relevant documents, the InformationRetrievalEvaluator computes different IR
## metrices. For our use case MRR@k and Accuracy@k are relevant.
#ir_evaluator = evaluation.InformationRetrievalEvaluator(ir_queries, ir_corpus, ir_relevant_docs)
#
#evaluators.append(ir_evaluator)
#
## Create a SequentialEvaluator. This SequentialEvaluator runs all three evaluators in a sequential order.
## We optimize the model with respect to the score from the last evaluator (scores[-1])
#seq_evaluator = evaluation.SequentialEvaluator(evaluators, main_score_function=lambda scores: scores[-1])
#
#
#logger.info("Evaluate model without training")
#seq_evaluator(model, epoch=0, steps=0, output_path=model_save_path)


# Train the model
model.fit(train_objectives=[(train_dataloader, train_loss)],
          evaluator=binary_acc_evaluator,
          epochs=num_epochs,
          warmup_steps=5,
          output_path=model_save_path
          )