train_val_test_split.py

# -*- coding: utf-8 -*-

import os
import sys
import json
import random
import argparse
import xml.etree.ElementTree as ET

def load_xml_file(path):
    """ Load an xml file.

    Args:
        path (:obj:`float`): Path where the XML file is stored.

    Returns:
        tree (:obj:`ElementTree`): Parsed tree structure of the XML file.


    """
    
    tree = ET.parse(path)

    return tree


def save_json(data, path):
    """ Save data as json file.

    Args:
        data (:obj:`list`): Data to store as json.
        path (:obj:`str`): Path where data will be stored.

    Returns:
        None

    """
    
    json.dump(data, open(path, 'w+', encoding='utf-8'))


def get_types(xml):
    """ Extract the different transformation types.

    Args:
        xml (:obj:`ElementTree`): XML as tree.

    Returns:
        transformations_types (:obj:`dict`): Mapping of transformations ids to their type

    Example:
        >>assert transformations_types[34] == 'Transformation - Noun to Verb' # True
        

    """
    
    transformations_types = {}
    root = xml.getroot()
    types = root[0]
    for curr_type in types:
        transformations_types[int(curr_type.attrib['id'])] = curr_type.text

    return transformations_types


def create_pair(text, simplified_text, transformation_id, transformation_type, source_dataset):
        """ Instanciate pair given parameters.

        Args:
            text (:obj:`str`): Raw text.
            simplified_text (:obj:`str`): Simplified text.
            transformation_id (:obj:`int`): Transformation ID
            transformation_type (:obj:`str`): Transformation Category
            source_dataset (:obj:`str`): Source dataset from which the pair comes from.

        Returns:
            pair (:obj:`dict`): Complex-Simplified Text pair with corresponding information.


        """
        pair = {}
        pair['transformation_id'] = transformation_id
        pair['transformation_type'] = transformation_type
        pair['source_dataset'] = source_dataset
        pair['text'] = text
        pair['simplified_text'] = simplified_text
        return pair


def fill_pairs_by_transformation(pairs_by_transformation, pair):
    """ This function adds a pair to the pairs_by_transformation dict. (into corresponding field).

    Args:
        pairs_by_transformation (:obj:`dict`): Dictionnary where pairs are organized by their
                                                transformation types.

        pair (:obj:`dict`): Complex-Simplified Text pair with corresponding information.

    Returns:
        None
    

    """

    transformation_id = pair['transformation_id']
    if transformation_id not in pairs_by_transformation.keys():
        pairs_by_transformation[transformation_id] = [pair]
    else:
        pairs_by_transformation[transformation_id].append(pair)


def fill_pairs_by_source_dataset(pairs_by_source_dataset, pair):
    """ This function adds a pair to the pairs_by_source_dataset dict. to corresponding field.

    Args:
        pairs_by_source_dataset (:obj:`dict`):  Dictionnary where pairs are organized by their
                                                source dataset.

        pair (:obj:`dict`): Complex-Simplified Text pair with corresponding information.

    Returns:
        None
    

    """

    source_dataset = pair['source_dataset']
    
    if source_dataset not in pairs_by_source_dataset.keys():
        pairs_by_source_dataset[source_dataset] = [pair]
    else:
        pairs_by_source_dataset[source_dataset].append(pair)


def get_pairs(xml, types):
    """ This function returns Complex-Simplified pairs from XML.

    Args:
        xml (:obj:`ElementTree`): XML simplifications pairs as tree structure.
        types (:obj:`dict`): Mapping of transformations ID to their transformations category.


    Returns:
        pairs (:obj:`list`): List of pairs (:obj:`dict`) without any ordering.
        pairs_by_transformation (:obj:`dict`): Pairs clustered by their transformation type.
        pairs_by_source_dataset (:obj:`dict`): Pairs clustered by their source dataset.

    """
    root = xml.getroot()
    simplifications = root[1]
    
    pairs = []
    pairs_by_transformation = {}
    pairs_by_source_dataset = {}
    for simplification in simplifications:
        transformation_id = int(simplification.attrib['type'])
        source = simplification.attrib['origin']
        raw_text = simplification[0].text
        simplified_text = simplification[1].text
        
        curr_pair = create_pair(raw_text, simplified_text, transformation_id, types[transformation_id], source)

        pairs.append(curr_pair)
        
        fill_pairs_by_transformation(pairs_by_transformation, curr_pair)
        fill_pairs_by_source_dataset(pairs_by_source_dataset, curr_pair)



    return pairs, pairs_by_transformation, pairs_by_source_dataset


def random_split(pairs, training_ratio):
    """ This function randomly splits pairs as train/val/test subsets.

    Args:
        pairs (:obj:`list`): List of pairs (:obj:`dict`) without any ordering.
        training_ratio (:obj:`float`): Ratio of training data (0<training_ratio<1).

    Returns:
        train (:obj:`list`): Training set
        validation (:obj:`list`): Validation set
        test (:obj:`list`): Testing set

    """

    random.shuffle(pairs)
    size = len(pairs)
    train_limit = int(size * training_ratio)
    val_limit = train_limit + int(size*(1-training_ratio)/2)
     
    train = pairs[:train_limit]
    val = pairs[train_limit:val_limit]
    test = pairs[val_limit:]

    return train, val, test



def challenge_seen_unseen_transformation_split(pairs_by_transformation, training_ratio):
    """ This function splits pairs s.t. evaluation can be done on seen and unseen
        transformations for more challenging robustness/generalization evaluation.
    
    Args:
        pairs_by_transformation (:obj:`dict`): Pairs organized by their transformation ID.
        training_ratio (:obj:`float`): Ratio of training data (0<training_ratio<1).

    Returns:
        train (:obj:`list`): Training set
        validation (:obj:`list`): Validation set
        seen_transformations_test (:obj:`list`): Seen Transformations testing set
        unseen_transformations_test (:obj:`list`): Unseen transformations testing set
    
    """

    # TODO transformations are hard-coded for now --> add argument in parser to specify them.
    seen_transformations_ids = [1, 2, 3, 11, 13, 23, 31, 32, 33, 34, 37]
    unseen_transformations_ids = [12, 22, 32, 35, 36]
    
    train = []
    val = []
    seen_transformations_test = []
    unseen_transformations_test = []

    for transf_id in seen_transformations_ids:
        curr_len = len(pairs_by_transformation[transf_id])
        train_limit = int(curr_len * training_ratio)
        val_limit = train_limit + int(curr_len * (1-training_ratio)/2.0)
        
        train += pairs_by_transformation[transf_id][:train_limit]
        val += pairs_by_transformation[transf_id][train_limit:val_limit]
        seen_transformations_test += pairs_by_transformation[transf_id][val_limit:]

    
    for transf_id in unseen_transformations_ids:
        unseen_transformations_test += pairs_by_transformation[transf_id]


    return train, val, seen_transformations_test, unseen_transformations_test


def challenge_seen_unseen_source_dataset_split(pairs_by_source_dataset, training_ratio):
    """ This function splits pairs s.t. evaluation can be done on seen and unseen
        source dataset for more challenging robustness/generalization evaluation.
    
    Args:
        pairs_by_source_dataset (:obj:`dict`): Pairs organized by their source dataset.
        training_ratio (:obj:`float`): Ratio of training data (0<training_ratio<1).

    Returns:
        train (:obj:`list`): Training set
        validation (:obj:`list`): Validation set
        seen_source_test (:obj:`list`): Seen source dataset testing set
        unseen_source_test (:obj:`list`): Unseen source dataset testing set
    
    """

    # TODO source dataset for training hard-coded --> add argument in parser.
    seen_source = ['itwiki']    # semi-supervised
    unseen_source = ['tn']  # manually annotated

    train = []
    val = []
    seen_source_test = []
    unseen_source_test = []

    for source in seen_source:
        random.shuffle(pairs_by_source_dataset[source])
        curr_len = len(pairs_by_source_dataset[source])
        train_limit = int(curr_len * training_ratio)
        val_limit = train_limit + int(curr_len * (1-training_ratio)/2.0)

        train += pairs_by_source_dataset[source][:train_limit]
        val += pairs_by_source_dataset[source][train_limit:val_limit]
        seen_source_test += pairs_by_source_dataset[source][val_limit:]

    for source in unseen_source:
        unseen_source_test += pairs_by_source_dataset[source]

    return train, val, seen_source_test, unseen_source_test


def split(args):
    """This function splits the XML file to produce specified subsets according to args.

    Args:
        args (:obj:`argparse.Namespace`): Parsed arguments.


    Returns:
        None

    """
    xml = load_xml_file(args.data_path)
    version = args.data_path[-6:-4]

    transformations_types = get_types(xml)
    pairs, pairs_by_transformation, pairs_by_source_dataset = get_pairs(xml, transformations_types)

    if args.split_criteria == 'random':
        train, val, test = random_split(pairs, args.training_ratio)
        os.makedirs(f'{args.out_dir}/{version}/random_split/', exist_ok=True)
        save_json(train, f'{args.out_dir}/{version}/random_split/train.json')
        save_json(val, f'{args.out_dir}/{version}/random_split/val.json')
        save_json(test, f'{args.out_dir}/{version}/random_split/test.json')

    elif args.split_criteria == 'transformations':
        seen_transformations_train, seen_transformations_val, seen_transformations_test, unseen_transformations_test = challenge_seen_unseen_transformation_split(pairs_by_transformation, args.training_ratio)
        
        os.makedirs(f'{args.out_dir}/{version}/transformations_split/', exist_ok=True)
        save_json(seen_transformations_train, f'{args.out_dir}/{version}/transformations_split/train.json')
        save_json(seen_transformations_val, f'{args.out_dir}/{version}/transformations_split/val.json')
        save_json(seen_transformations_test, f'{args.out_dir}/{version}/transformations_split/seen_transformations_test.json')
        save_json(unseen_transformations_test, f'{args.out_dir}/{version}/transformations_split/unseen_transformations_test.json')

    elif args.split_criteria == 'source_dataset':
        
        itwiki_train, itwiki_val, itwiki_test, tn_test = challenge_seen_unseen_source_dataset_split(pairs_by_source_dataset, args.training_ratio)
        
        os.makedirs(f'{args.out_dir}/{version}/source_dataset_split/', exist_ok=True)
        save_json(itwiki_train, f'{args.out_dir}/{version}/source_dataset_split/itwiki_train.json')
        save_json(itwiki_val, f'{args.out_dir}/{version}/source_dataset_split/itwiki_val.json')
        save_json(itwiki_test, f'{args.out_dir}/{version}/source_dataset_split/itwiki_test.json')
        save_json(tn_test, f'{args.out_dir}/{version}/source_dataset_split/tn_test.json')
        


if __name__ == '__main__':

    parser = argparse.ArgumentParser(description="Creating Train/Val/Test files")

    parser.add_argument('--data_path', type=str, required=True, help='path to (single) data file')
    parser.add_argument('--out_dir', type=str, required=True, help='output dir to store files')

    parser.add_argument('--training_ratio', type=float, required=True, help='training ratio (e.g. 0.8). Remaining will be divided for val and test EQUALLY.')

    parser.add_argument('--split_criteria', type=str, required=True, choices=['random', 'transformations', 'source_dataset'], help='split criteria')

    
    args = parser.parse_args()
    split(args)
    
    sys.exit(0)