BERTでKaggleの過去問Quora Question Pairsを解いてみる

現在作成中です．．．もう少々お待ち下さい．．．

タスク

• https://www.kaggle.com/c/quora-question-pairs
• 質問の重複一致判定コンペ

開発環境

• GPU: GTX-1080 Ti(多分，メモリが10GB以上必要)
• TPU: 利用してない
• 動作しない場合はtrain_batch_sizeの値を小さくしてください．
• BERT Model: (Base版でLarge版ではないです)

前準備

• コンペのデータセットのダウンロード
• bertのコードと学習済みモデルのダウンロード

mkdir input
kaggle competitions download -c quora-question-pairs -p input
unzip "input/*.zip" -d input
git clone https://github.com/google-research/bert.git
wget https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-12_H-768_A-12.zip
unzip uncased_L-12_H-768_A-12.zip

コード

# coding: utf-8
import os
import sys
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.model_selection import train_test_split

sys.path.append("./bert")
import modeling
import optimization
import run_classifier
import tokenization

DEBUG_MODE = False
config = {
    "qqp_model_directory": "/tmp/bert_model",
    "bert_config_filepath": "./uncased_L-12_H-768_A-12/bert_config.json",
    "vocab_filepth": "./uncased_L-12_H-768_A-12/vocab.txt",
    "input_directory": "./input",
    "data_directory": "./data",
    "output_directory": "./output",
    "init_checkpoint": None,
    "do_lower_case": True,
    "max_seq_length": 128,
    "train_batch_size": 32,
    "eval_batch_size": 8,
    "predict_batch_size": 8,
    "learning_rate": 5e-5,
    "num_train_epochs": 3.0,
    "warmup_proportion": 0.1,
    "save_checkpoints_steps": 1000,
    "nrows": (2000 if DEBUG_MODE else None),
}

for directory in [config["data_directory"], config["output_directory"], config["qqp_model_directory"]]:
    if not os.path.exists(directory):
        os.makedirs(directory)

# 壊れたデータの修正

train_df = pd.read_csv(os.path.join(config["input_directory"], "train.csv"), encoding="utf-8")
test_df = pd.read_csv(os.path.join(config["input_directory"], "test.csv"), encoding="utf-8")

def remove(df, on):
    # cf. Test data: https://www.kaggle.com/c/quora-question-pairs/discussion/59035
    valid_ids =[type(x)==int for x in df[on]]
    return df[valid_ids].drop_duplicates()

def fillna(df, value=""):
    df["question1"].fillna(value, inplace=True)
    df["question2"].fillna(value, inplace=True)
    return df

train_df = fillna(remove(train_df, on="id"), value="__NA__")
test_df = fillna(remove(test_df, on="test_id"), value="__NA__")

train_df.to_csv(os.path.join(config["data_directory"], "train.csv"), encoding="utf-8", index=False)
test_df.to_csv(os.path.join(config["data_directory"], "test.csv"), encoding="utf-8", index=False)

# Bert用に変換

class QqpProcessor(object): # QQP: Quora Question Pairs
    def __init__(self, data_directory, nrows=None, random_state=777):
        self._all_train_df = pd.read_csv(os.path.join(data_directory, "train.csv"), nrows=nrows, encoding="utf-8")
        self._train_df, self._dev_df = train_test_split(self._all_train_df, random_state=777)
        self._test_df = pd.read_csv(os.path.join(data_directory, "test.csv"), nrows=nrows, encoding="utf-8")
        self._test_df["id"] = self._test_df["test_id"]

    def get_train_examples(self, data_dir):
        return self._create_examples(self._train_df, "train")

    def get_dev_examples(self, data_dir):
        return self._create_examples(self._dev_df, "dev")

    def get_test_examples(self, data_dir):
        return self._create_examples(self._test_df, "test")

    def get_labels(self):
        return ["0", "1"]

    def _create_examples(self, df, set_type):
        def create_example(row):
            guid = row["id"]
            text_a = tokenization.convert_to_unicode(row["question1"])
            text_b = tokenization.convert_to_unicode(row["question2"])
            if set_type == "test":
                label = "0"
            else: # train, dev
                label = tokenization.convert_to_unicode(str(row["is_duplicate"]))
            return run_classifier.InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)
        return df.apply(create_example, axis=1).values.tolist()

processor = QqpProcessor(data_directory=config["data_directory"], nrows=config["nrows"])
tokenization.validate_case_matches_checkpoint(do_lower_case=config["do_lower_case"], init_checkpoint=None)
bert_config = modeling.BertConfig.from_json_file(config["bert_config_filepath"])
label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer(vocab_file=config["vocab_filepth"], do_lower_case=config["do_lower_case"])
run_config = tf.contrib.tpu.RunConfig(
    model_dir=config["qqp_model_directory"],
    save_checkpoints_steps=1000,
)
train_examples = processor.get_train_examples(config["data_directory"])
num_train_steps = int(len(train_examples) / config["train_batch_size"] * config["num_train_epochs"])
num_warmup_steps = int(num_train_steps * config["warmup_proportion"])

model_fn = run_classifier.model_fn_builder(
    bert_config=bert_config,
    num_labels=len(label_list),
    init_checkpoint=config["init_checkpoint"],
    learning_rate=config["learning_rate"],
    num_train_steps=num_train_steps,
    num_warmup_steps=num_warmup_steps,
    use_tpu=False,
    use_one_hot_embeddings=False
)

estimator = tf.contrib.tpu.TPUEstimator( # TPU使わない場合自動的にGPUかCPUに切り替わる
    use_tpu=False,
    model_fn=model_fn,
    config=run_config,
    train_batch_size=config["train_batch_size"],
    eval_batch_size=config["eval_batch_size"],
    predict_batch_size=config["predict_batch_size"]
)

# Training

train_file = os.path.join(config["output_directory"], "train.tf_record")
run_classifier.file_based_convert_examples_to_features(train_examples, label_list, config["max_seq_length"], tokenizer, train_file)
train_input_fn = run_classifier.file_based_input_fn_builder(
    input_file=train_file,
    seq_length=config["max_seq_length"],
    is_training=True,
    drop_remainder=True
)
estimator.train(input_fn=train_input_fn, max_steps=num_train_steps)

# Prediction

predict_examples = processor.get_test_examples(config["data_directory"])
num_actual_predict_examples = len(predict_examples)
predict_file = os.path.join(config["output_directory"], "predict.tf_record")
run_classifier.file_based_convert_examples_to_features(
    predict_examples, 
    label_list,                                    
    config["max_seq_length"], 
    tokenizer,                                    
    predict_file
)
predict_drop_remainder = False
predict_input_fn = run_classifier.file_based_input_fn_builder(
    input_file=predict_file,
    seq_length=config["max_seq_length"],
    is_training=False,
    drop_remainder=predict_drop_remainder
)

result = estimator.predict(input_fn=predict_input_fn)
probabilities = np.array([prediction["probabilities"][1] for (i, prediction) in enumerate(result)])
processor._test_df["is_duplicate"] = probabilities
processor._test_df[["is_duplicate", "test_id"]].to_csv("submission.csv", index=False)
pd.read_csv("submission.csv").head()

結果

このコンペは同じ質問が何度も使われており，使われる頻度が重要なヒントになっています．
そのため，BERT単体では，

• 学習時間: 3〜4時間
• 予測時間: 3時間
• Private: 0.33466
• Public: 0.32676

となり，あまり良い性能は出ない状況です．

見た感じ，Private: 0.12988となっている24位の解法がシンプルで，LSTM部分をBERTに置き換えてグラフ構造的な特徴量を組み込んだネットワークを作れば良さそう

まとめ

• 質問の利用頻度などグラフ構造から得られる特徴量が重要なので，BERT単体では性能は出せない模様
• extract_features.py形式で特徴量が出せるので，特徴を抽出してネットワークを構築する必要あり．(もう少々お待ち下さい...)
• 24位の解法がシンプルなので，これベースに作っていったら良さそうです．