utils.py

# coding: utf-8
# Author: Miracle Yoo
import codecs
import os
import datetime
import numpy as np
import pandas as pd
import pickle

import warnings

warnings.filterwarnings("ignore")


def folder_init():
    """
    Initialize folders required
    """
    if not os.path.exists('source'): os.mkdir('source')
    if not os.path.exists('source/log'): os.mkdir('source/log')
    if not os.path.exists('source/trained_net'): os.mkdir('source/trained_net')
    if not os.path.exists('source/data'): os.mkdir('source/data')
    if not os.path.exists('source/nohup'): os.mkdir('source/nohup')
    if not os.path.exists('source/summaries'): os.mkdir('source/summaries')


def change_dataset_format(filename):
    """
    Change the csv dataset file which fanyu using to stadard || dataset file.
    :param filename: dataset filename
    """
    oridf = pd.read_csv(filename)
    data = []
    for line in oridf.iterrows():
        if type(line[1]['query']) != float and type(line[1]['main_question']) != float:
            data.append(line[1]['query'] + '||' + line[1]['main_question'] + '\n')
    with open(os.path.splitext(filename)[0] + '.txt', 'w+') as f:
        f.writelines(data)


def multi_dataset_merge(*filenames):
    """
    :param filenames: multiple dataset path
    :return: merged dataset list
    """
    data = []
    for filename in filenames:
        with open(filename) as f:
            data.extend(f.readlines())
    return data


def get_equal_pairs(title):
    pairs = []
    with codecs.open('./reference/equal_pairs.txt', 'r', encoding='utf-8') as f:
        raw_pairs = f.readlines()
    for line in raw_pairs:
        pairs.append([title.index(x.strip('\n')) for x in line.split('||') if x.strip('\n') in title])
    return pairs


def use_pairs_mapping(predicts, labels, equal_pairs):
    """
    :param predicts:
    :param labels:
    :param equal_pairs:
    :return:
    """
    correct_num = 0
    # 如果输入的predict是一个[batch_size,top_n_num]的矩阵，
    # 用于预测top-n准确率
    if type(predicts[0]) != int:
        for i in range(len(labels)):
            if labels[i] in predicts[i]:
                correct_num += 1
            else:
                for j in range(len(equal_pairs)):
                    if labels[i] in equal_pairs[j]:
                        for pred in predicts[i]:
                            if pred in equal_pairs[j]:
                                correct_num += 1
                                break
    # 如果输入的predict是一个[batch_size]的list，
    # 用于预测top-1准确率
    else:
        for i in range(len(labels)):
            if labels[i] == predicts[i]:
                correct_num += 1
            else:
                for j in range(len(equal_pairs)):
                    if labels[i] in equal_pairs[j] and predicts[i] in equal_pairs[j]:
                        correct_num += 1
    return correct_num


def get_topn_acc(outputs, labels, top_num=3):
    acc = [0] * top_num
    for i in range(top_num):
        predicts = np.array(outputs.sort(descending=True, dim=1)[1])[:, :top_num]
        for j in range(len(labels)):
            if labels[j] in predicts:
                acc[i] += 1
    acc = [x / len(labels) for x in acc]
    return acc


def write_summary(net, opt, summary_info):
    print(summary_info['best_top_accs'])
    current_time = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
    prefix = './source/summaries/' + net.model_name
    if not os.path.exists(prefix): os.mkdir(prefix)
    sum_path = prefix + '/MiracleYoo_' + current_time + '_' + net.model_name + '_Model_Record_Form.md'
    with codecs.open('./config.py', 'r', encoding='utf-8') as f:
        raw_data = f.readlines()
        configs = ''
        for line in raw_data:
            if line.strip().startswith('self.'):
                configs += line.strip().lstrip('self.') + '\n'

    content = '''
        # Model Testing Record Form
        | Item Name        | Information |
        | ---------        | ----------- |
        | Model Name       | %s          |
        | Tester's Name    | Miracle Yoo |
        | Author's Nmae    | Miracle Yoo |
        | Test Time        | %s          |
        | Test Position    | %s          |
        | Training Epoch   | %d          |
        | Highest Test Acc | %.4f        |
        | Loss of best Test Acc | %.4f   |
        | Top2Acc of best Test Acc|%.4f  |
        | Top3Acc of best Test Acc|%.4f  |
        | Last epoch test acc   | %.4f   |
        | Last epoch test loss  | %.4f   |
        | Last epoch train acc  | %.4f   |
        | Last epoch train loss | %.4f   |
        | Train Dataset Path    | %s     |
        | Test Dataset Path     | %s     |
        | Class Number     | %d          |
        | Framwork         | Pytorch     |
        | Basic Method     | Classify    |
        | Input Type       | Char        |
        | Criterion        | CrossEntropy|
        | Optimizer        | %s          |
        | Learning Rate    | %.4f        |
        | Embedding dimension   | %d     |
        | Data Homogenization   | True   |
        | Pretreatment|Remove punctuation|
        | Other Major Param |            |
        | Other Operation   |            |
        
        
        ## Configs
        ```
        %s
        ```
        
        ## Net Structure
        ```
        %s
        ```
    ''' % (
        net.model_name,
        current_time,
        opt.TEST_POSITION,
        summary_info['total_epoch'],
        summary_info['best_acc'],
        summary_info['best_acc_loss'],
        summary_info['best_top_accs'][1],
        summary_info['best_top_accs'][2],
        summary_info['ave_test_acc'],
        summary_info['ave_test_loss'],
        summary_info['ave_train_acc'],
        summary_info['ave_train_loss'],
        os.path.basename(opt.TRAIN_DATASET_PATH),
        os.path.basename(opt.TEST_DATASET_PATH),
        opt.NUM_CLASSES,
        opt.OPTIMIZER,
        opt.LEARNING_RATE,
        opt.EMBEDDING_DIM,
        configs.strip('\n'),
        str(net)
    )
    with codecs.open(sum_path, 'w+', encoding='utf-8') as f:
        f.writelines(content)