md_bayesian_er_ditto/ml_er/ditto_er.py

import os
import sys
import time

os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
sys.path.append('/root/hjt/md_bayesian_er_ditto/')
import pdb
import pickle
import torch
import json
import numpy as np
import pandas as pd
import random
from tqdm import tqdm
from setting import *
from colorama import Fore
from argparse import Namespace
import ConfigSpace
from ConfigSpace import Configuration
from ditto.matcher import set_seed, predict, tune_threshold, load_model
from ConfigSpace.read_and_write import json as csj
from ditto.ditto_light.dataset import DittoDataset
from ditto.ditto_light.summarize import Summarizer
from ditto.ditto_light.knowledge import *
from ditto.ditto_light.ditto import train


def matching(hpo_config):
    print(Fore.BLUE + f'Config: {hpo_config}')

    with open(md_output_dir + "/mds.pickle", "rb") as file:
        md_list = pickle.load(file)

    hp = Namespace()
    hp.task = directory_path.replace('/root/hjt/DeepMatcher Dataset/', '')

    # only a single task for baseline
    task = hp.task
    # load task configuration
    configs = json.load(open('../ditto/configs.json'))
    configs = {conf['name']: conf for conf in configs}
    config = configs[task]

    config['trainset'] = '/root/hjt/md_bayesian_er_ditto/ditto/' + config['trainset']
    config['validset'] = '/root/hjt/md_bayesian_er_ditto/ditto/' + config['validset']
    config['testset'] = '/root/hjt/md_bayesian_er_ditto/ditto/' + config['testset']

    trainset = config['trainset']
    validset = config['validset']
    testset = config['testset']

    hp.run_id = 0
    hp.batch_size = hpo_config['batch_size']
    hp.max_len = hpo_config['max_len']
    hp.lr = 3e-5
    # hp.n_epochs = 20
    hp.n_epochs = 2
    # hp.finetuning
    hp.save_model = True
    hp.input_path = config['testset']
    hp.output_path = '/root/autodl-tmp/output/predictions.jsonl'

    hp.logdir = '/root/autodl-tmp/checkpoints/'
    hp.checkpoint_path = '/root/autodl-tmp/checkpoints/'
    hp.lm = hpo_config['language_model']
    hp.fp16 = hpo_config['half_precision_float']
    hp.da = hpo_config['data_augmentation']
    hp.alpha_aug = 0.8
    hp.dk = None
    hp.summarize = hpo_config['summarize']
    hp.size = None
    hp.use_gpu = True

    seed = hp.run_id
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)

    # create the tag of the run
    run_tag = '%s_lm=%s_da=%s_dk=%s_su=%s_size=%s_id=%d' % (task, hp.lm, hp.da,
                                                            hp.dk, hp.summarize, str(hp.size), hp.run_id)
    run_tag = run_tag.replace('/', '_')

    # summarize the sequences up to the max sequence length
    if hp.summarize:
        summarizer = Summarizer(config, lm=hp.lm)
        trainset = summarizer.transform_file(trainset, max_len=hp.max_len)
        validset = summarizer.transform_file(validset, max_len=hp.max_len)
        testset = summarizer.transform_file(testset, max_len=hp.max_len)

    # load train/dev/test sets
    train_dataset = DittoDataset(trainset,
                                 lm=hp.lm,
                                 max_len=hp.max_len,
                                 size=hp.size,
                                 da=hp.da)
    valid_dataset = DittoDataset(validset, lm=hp.lm)
    test_dataset = DittoDataset(testset, lm=hp.lm)

    # train and evaluate the model
    train(train_dataset,
          valid_dataset,
          test_dataset,
          run_tag, hp)

    set_seed(123)
    config, model = load_model(hp.task, hp.checkpoint_path,
                               hp.lm, hp.use_gpu, hp.fp16)

    summarizer = dk_injector = None

    if hp.summarize:
        summarizer = Summarizer(config, hp.lm)

    # tune threshold
    threshold = tune_threshold(config, model, hp)

    # run prediction
    predict(hp.input_path, hp.output_path, config, model,
            summarizer=summarizer,
            max_len=hp.max_len,
            lm=hp.lm,
            dk_injector=dk_injector,
            threshold=threshold)

    predictions_raw = pd.read_json(hp.output_path, encoding='ISO-8859-1', lines=True)
    predictions = pd.read_csv(directory_path + '/test_whole.csv', encoding='ISO-8859-1')
    predictions['predicted'] = predictions_raw['match']
    indicators = evaluate_prediction(predictions, 'label', 'predicted')
    predictions.drop(columns='_id', inplace=True)
    predictions = predictions.reset_index(drop=True)
    predictions = predictions.astype(str)

    sim_tensor_dict = build_col_pairs_sim_tensor_dict(predictions)
    predictions['confidence'] = 0
    predictions['md'] = ''

    epl_match = 0  # 可解释，预测match
    if len(md_list) > 0:
        for row in tqdm(predictions.itertuples()):
            if str(getattr(row, 'predicted')) == str(1):
                conf, md_dict = is_explicable(row, md_list, sim_tensor_dict)
                if conf > 0:
                    predictions.loc[row[0], 'confidence'] = conf
                    predictions.loc[row[0], 'md'] = str(md_dict)
                    epl_match += 1

    df = predictions[predictions['predicted'] == str(1)]
    interpretability = epl_match / len(df)  # 可解释性
    indicators['interpretability'] = interpretability
    performance = interpre_weight * interpretability + (1 - interpre_weight) * indicators["F1"]
    indicators['performance'] = performance
    print(Fore.BLUE + f'ER Indicators: {indicators}')
    predictions.to_csv(er_output_dir + '/predictions.csv', sep=',', index=False, header=True)
    print(Fore.CYAN + f'Finish Time: {time.time()}')
    return indicators


def evaluate_prediction(prediction_: pd.DataFrame, labeled_attr: str, predicted_attr: str) -> dict:
    new_df = prediction_.reset_index(drop=False, inplace=False)
    gold = new_df[labeled_attr]
    predicted = new_df[predicted_attr]
    gold_negative = gold[gold == 0].index.values
    gold_positive = gold[gold == 1].index.values
    predicted_negative = predicted[predicted == 0].index.values
    predicted_positive = predicted[predicted == 1].index.values

    false_positive_indices = list(set(gold_negative).intersection(predicted_positive))
    true_positive_indices = list(set(gold_positive).intersection(predicted_positive))
    false_negative_indices = list(set(gold_positive).intersection(predicted_negative))

    num_true_positives = float(len(true_positive_indices))
    num_false_positives = float(len(false_positive_indices))
    num_false_negatives = float(len(false_negative_indices))

    precision_denominator = num_true_positives + num_false_positives
    recall_denominator = num_true_positives + num_false_negatives

    precision = 0.0 if precision_denominator == 0.0 else num_true_positives / precision_denominator
    recall = 0.0 if recall_denominator == 0.0 else num_true_positives / recall_denominator
    F1 = 0.0 if precision == 0.0 and recall == 0.0 else (2.0 * precision * recall) / (precision + recall)

    return {"precision": precision, "recall": recall, "F1": F1}


def build_col_pairs_sim_tensor_dict(predictions: pd.DataFrame):
    predictions_attrs = predictions.columns.values.tolist()
    col_tuple_list = []
    for _ in predictions_attrs:
        if _.startswith('ltable'):
            left_index = predictions_attrs.index(_)
            right_index = predictions_attrs.index(_.replace('ltable_', 'rtable_'))
            col_tuple_list.append((left_index, right_index))

    length = predictions.shape[0]
    # width = predictions.shape[1]
    predictions = predictions.reset_index(drop=True)
    sentences = predictions.values.flatten(order='F').tolist()

    embedding = model.encode(sentences, convert_to_tensor=True, device="cuda", batch_size=256, show_progress_bar=True)
    split_embedding = torch.split(embedding, length, dim=0)
    table_tensor = torch.stack(split_embedding, dim=0, out=None)
    # prediction的归一化嵌入张量
    norm_table_tensor = torch.nn.functional.normalize(table_tensor, dim=2)
    sim_tensor_dict = {}
    for col_tuple in col_tuple_list:
        lattr_tensor = norm_table_tensor[col_tuple[0]]
        rattr_tensor = norm_table_tensor[col_tuple[1]]
        mul_tensor = lattr_tensor * rattr_tensor
        sim_tensor = torch.sum(mul_tensor, 1)
        sim_tensor = torch.round(sim_tensor * 100) / 100
        sim_tensor_dict[predictions_attrs[col_tuple[0]].replace('ltable_', '')] = sim_tensor
    return sim_tensor_dict


def is_explicable(row, all_mds: list, st_dict):
    attrs = all_mds[0][0].keys()  # 从第一条md_tuple中的md字典中读取所有字段
    for md_tuple in all_mds:
        explicable = True  # 假设这条md能解释当前元组
        for a in attrs:
            if st_dict[a][row[0]].item() < md_tuple[0][a]:
                explicable = False  # 任意一个字段的相似度达不到阈值，这条md就不能解释当前元组
                break  # 不再与当前md的其他相似度阈值比较，跳转到下一条md
        if explicable:
            return md_tuple[2], md_tuple[0]  # 任意一条md能解释，直接返回
    return -1.0, {}  # 遍历结束，不能解释


def ml_er(config: Configuration):
    indicators = matching(config)
    output_path = er_output_dir + "/eval_result.txt"
    with open(output_path, 'w') as _f:
        _f.write('precision:' + str(indicators['precision']) + '\n')
        _f.write('recall:' + str(indicators['recall']) + '\n')
        _f.write('F1:' + str(indicators["F1"]) + '\n')
        _f.write('interpretability:' + str(indicators['interpretability']) + '\n')
        _f.write('performance:' + str(indicators['performance']) + '\n')


if __name__ == '__main__':
    if os.path.isfile(hpo_output_dir + "/incumbent.json"):
        with open(hpo_output_dir + "/configspace.json", 'r') as f:
            dict_configspace = json.load(f)
        str_configspace = json.dumps(dict_configspace)
        configspace = csj.read(str_configspace)
        with open(hpo_output_dir + "/incumbent.json", 'r') as f:
            dic = json.load(f)
        configuration = ConfigSpace.Configuration(configspace, values=dic)
        ml_er(configuration)