train_valid_split.py

import os
import json
import tqdm
import functools
import collections
import multiprocessing
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import linear_kernel


def extract_domains(filename):
    domains = set()
    with open(filename) as f:
        for line in f:
            line = json.loads(line.strip())
            domains.add(line["domain"])
    return filename, list(domains)


def filter_valid(questions):
    answers = set()
    new_questions = []
    for question in questions:
        if question["answer"] not in answers:
            new_questions.append(question)
        answers.add(question["answer"])
    return new_questions


# def format_to_valid(questions):
#     answers_txt = [e["answer"] for e in questions]
#     questions_txt = [e["question"] for e in questions]
#     vectorizer = TfidfVectorizer()
#     vectorizer.fit(answers_txt + questions_txt)
#     answer_vectors = vectorizer.transform(answers_txt)
#     for i, question in enumerate(questions):
#         similarities = linear_kernel(answer_vectors[[i]], answer_vectors).flatten()
#         answer_scores = [(j, sim) for j, sim in enumerate(similarities) if sim != 1]
#         answer_scores = sorted(answer_scores, key=lambda x: x[1], reverse=True)
#         sorted_answers = [questions[j]["answer"] for j, _ in answer_scores if questions[j]["answer"] != question["answer"]]
#         negative_answer = sorted_answers[len(sorted_answers) // 2]
#         assert question["answer"] not in sorted_answers
#         question["candidates"] = [question["answer"]] + sorted_answers
#         question["negative_example"] = negative_answer
#     return questions


def format_to_valid(questions):
    answers = [e["answer"] for e in questions]
    for question in questions:
        answer = question["answer"]
        candidates = [e for e in answers if e != answer]
        candidates = [answer] + candidates
        question["candidates"] = candidates
    return questions


def format_to_train(questions):
    answers_txt = [e["answer"] for e in questions]
    answers_shifted = answers_txt[1:] + [answers_txt[0]]
    for question, answer in zip(questions, answers_shifted):
        question["negative"] = answer
    return questions


def valid_train_split(filename, mapping=None):
    previous_domain = ""
    train = []
    valid = []
    domain_data = {"questions": [], "pages": set()}
    counter = 0
    with open(filename) as f:
        for line_txt in f:
            counter += 1
            line = json.loads(line_txt.strip())
            domain = line["domain"]
            if domain != previous_domain and previous_domain != "":
                form_questions = format_to_train(domain_data["questions"])
                if len(mapping[previous_domain]) > 1:
                    train.extend(form_questions)
                elif len(valid) > 2000:
                    train.extend(form_questions)
                elif len(domain_data["pages"]) > 1:
                    train.extend(form_questions)
                elif len(domain_data["questions"]) < 15:
                    train.extend(form_questions)
                else:
                    questions = filter_valid(domain_data["questions"])
                    if len(questions) < 15:
                        train.extend(form_questions)
                    else:
                        questions = format_to_valid(questions)
                        valid.extend(questions)
                domain_data = {"questions": [], "pages": set()}
            domain_data["questions"].append(line)
            domain_data["pages"].add(line["domain_index"])
            previous_domain = domain
        # train.extend(form_questions)
    return train, valid, filename


domain_count = collections.defaultdict(list)
data = [f"data/{e}" for e in os.listdir("data") if e.endswith(".json")]
# with multiprocessing.Pool(os.cpu_count()) as p:
with multiprocessing.Pool(1) as p:
    for filename, domains in tqdm.tqdm(p.imap_unordered(extract_domains, data)):
        language = filename.split(".")[1]
        for domain in domains:
            domain_count[domain].append(language)


with multiprocessing.Pool(os.cpu_count()) as p:
    fn = functools.partial(valid_train_split, mapping=domain_count)
    for train, valid, filename in tqdm.tqdm(p.imap_unordered(fn, data)):
        train_filename = filename.replace("data/", "data/train/")
        train = [json.dumps(e, ensure_ascii=False) for e in train]
        valid = [json.dumps(e, ensure_ascii=False) for e in valid]
        with open(train_filename, "w+") as f:
            train = "\n".join(train)
            f.write(train) 
        valid_filename = filename.replace("data/", "data/valid/")
        with open(valid_filename, "w+") as f:
            valid = "\n".join(valid)
            f.write(valid)