preprocess.py

#!/usr/bin/env python
import onmt
import onmt.markdown
import argparse
import torch
import subprocess
import time, datetime
from onmt.data.binarizer import Binarizer
from onmt.data.binarizer import SpeechBinarizer

from onmt.data.indexed_dataset import IndexedDatasetBuilder

import numpy as np
import warnings
import os
from os.path import dirname, abspath
import gc



warnings.filterwarnings("ignore", category=UserWarning)

parser = argparse.ArgumentParser(description='preprocess.py')
onmt.markdown.add_md_help_argument(parser)

# **Preprocess Options**
parser.add_argument('-multi_dataset', action='store_true',
                    help="Save each dataset separately instead of one joined dataset")
parser.add_argument('-multi_mirror', action='store_true',
                    help="Save each dataset separately instead of one joined dataset")
parser.add_argument('-resume', action='store_true',
                    help="If the dataset is created, ignored and create the next one")
parser.add_argument('-config', help="Read options from this file")

parser.add_argument('-src_type', default="text",
                    help="Type of the source input. Options are [text|img|audio].")
parser.add_argument('-sort_type', default="ascending",
                    help="Type of sorting. Options are [ascending|descending].")
parser.add_argument('-src_img_dir', default=".",
                    help="Location of source images")
parser.add_argument('-stride', type=int, default=1,
                    help="Stride on input features")
parser.add_argument('-concat', type=int, default=1,
                    help="Concate sequential audio features to decrease sequence length")
parser.add_argument('-previous_context', type=int, default=0,
                    help="Number of previous sentence for context")
parser.add_argument('-input_type', default="word",
                    help="Input type: word/char")
parser.add_argument('-data_type', default="int64",
                    help="Input type for storing text (int64|int32|int|int16) to reduce memory load")
parser.add_argument('-format', default="raw",
                    help="Save data format: binary or raw. Binary should be used to load faster")
parser.add_argument('-external_tokenizer', default="",
                    help="External tokenizer from Huggingface. Currently supports barts.")

parser.add_argument('-train_src', required=True,
                    help="Path to the training source data")
parser.add_argument('-past_train_src', default="",
                    help="Path to the training source data")
parser.add_argument('-future_train_src', default="",
                    help="Path to the training source data")
parser.add_argument('-train_tgt', required=True,
                    help="Path to the training target data")
parser.add_argument('-valid_src', required=True,
                    help="Path to the validation source data")
parser.add_argument('-past_valid_src', default="",
                    help="Path to the validation source data")
parser.add_argument('-future_valid_src', default="",
                    help="Path to the validation source data")
parser.add_argument('-valid_tgt', required=True,
                    help="Path to the validation target data")

parser.add_argument('-train_src_lang', default="src",
                    help="Language(s) of the source sequences.")
parser.add_argument('-train_src_atbs', default="",
                    help="Attributes(s) of the source sequences.")

parser.add_argument('-train_tgt_lang', default="tgt",
                    help="Language(s) of the target sequences.")
parser.add_argument('-train_tgt_atbs', default="",
                    help="Attributes(s) of the source sequences.")

parser.add_argument('-valid_src_lang', default="src",
                    help="Language(s) of the source sequences.")
parser.add_argument('-valid_src_atbs', default="",
                    help="Attributes(s) of the source sequences.")

parser.add_argument('-valid_tgt_lang', default="tgt",
                    help="Language(s) of the target sequences.")
parser.add_argument('-valid_tgt_atbs', default="",
                    help="Attributes(s) of the source sequences.")

parser.add_argument('-save_data', required=True,
                    help="Output file for the prepared data")

parser.add_argument('-src_vocab_size', type=int, default=9999999,
                    help="Size of the source vocabulary")
parser.add_argument('-tgt_vocab_size', type=int, default=9999999,
                    help="Size of the target vocabulary")
parser.add_argument('-src_vocab',

                    help="Path to an existing source vocabulary")
parser.add_argument('-tgt_vocab',
                    help="Path to an existing target vocabulary")

parser.add_argument('-load_dict',
                    help="Path to an existing target vocabulary")

parser.add_argument('-src_seq_length', type=int, default=10000,
                    help="Maximum source sequence length")
parser.add_argument('-src_seq_length_trunc', type=int, default=0,
                    help="Truncate source sequence length.")
parser.add_argument('-tgt_seq_length', type=int, default=10000,
                    help="Maximum target sequence length to keep.")
parser.add_argument('-tgt_seq_length_trunc', type=int, default=0,
                    help="Truncate target sequence length.")

# tokens
parser.add_argument('-src_bos_token', type=str, default="<s>",
                    help='SRC BOS Token Default is <s>.')
parser.add_argument('-src_eos_token', type=str, default="</s>",
                    help='SRC BOS Token. Default is </s>.')
parser.add_argument('-src_unk_token', type=str, default="<unk>",
                    help='SRC Unk Token. Default is <unk>.')
parser.add_argument('-src_pad_token', type=str, default="<blank>",
                    help='SRC PAD Token. Default is <blank>.')

parser.add_argument('-tgt_bos_token', type=str, default="<s>",
                    help='TGT BOS Token Default is <s>.')
parser.add_argument('-tgt_eos_token', type=str, default="</s>",
                    help='TGT BOS Token. Default is </s>.')
parser.add_argument('-tgt_unk_token', type=str, default="<unk>",
                    help='TGT Unk Token. Default is <unk>.')
parser.add_argument('-tgt_pad_token', type=str, default="<blank>",
                    help='TGT PAD Token. Default is <blank>.')

parser.add_argument('-shuffle', type=int, default=1,
                    help="Shuffle data")

parser.add_argument('-asr', action='store_true',
                    help="prepare data for asr task")
parser.add_argument('-asr_format', default="h5",
                    help="Format of asr data h5 or scp")
parser.add_argument('-lm', action='store_true',
                    help="prepare data for LM task")
parser.add_argument('-fp16', action='store_true',
                    help="store ASR data in fp16")

parser.add_argument('-seed', type=int, default=3435,
                    help="Random seed")

parser.add_argument('-lower', action='store_true', help='lowercase data')
parser.add_argument('-load_bpe_voc', action='store_true', help='')
parser.add_argument('-no_bos', action='store_true', help='not adding bos word (this is done manually in the data)')
parser.add_argument('-sort_by_target', action='store_true', help='')
parser.add_argument('-join_vocab', action='store_true', help='Using one dictionary for both source and target')

parser.add_argument('-report_every', type=int, default=100000,
                    help="Report status every this many sentences")
parser.add_argument('-reshape_speech', type=int, default=1,
                    help="Reshaping the speech segments here. Mostly for compatibility..")
parser.add_argument('-num_threads', type=int, default=1,
                    help="Number of threads for multiprocessing")
parser.add_argument('-verbose', action='store_true',
                    help="Print out information during preprocessing")
parser.add_argument('-num_mel_bin', type=int, default=0,
                    help="Number of Log Mel bin, if > 0 the waveforms will be converted to LogMel features")


opt = parser.parse_args()
torch.manual_seed(opt.seed)


def make_vocab(name, filenames, size, tokenizer, num_workers=1):
    if name == "source":
        vocab = onmt.Dict([opt.src_pad_token, opt.src_unk_token,
                           opt.src_bos_token, opt.src_eos_token],
                          lower=opt.lower)
    elif name == "target":
        vocab = onmt.Dict([opt.tgt_pad_token, opt.tgt_unk_token,
                           opt.tgt_bos_token, opt.tgt_eos_token],
                          lower=opt.lower)
    else:
        print("Warning: check the name")
        exit(-1)

    for filename in filenames:
        print("Generating vocabulary from file %s ... " % filename)
        onmt.Dict.gen_dict_from_file(filename, vocab, tokenizer, num_workers=num_workers)

    original_size = vocab.size()
    vocab = vocab.prune(size)
    print('Created dictionary of size %d (pruned from %d)' %
          (vocab.size(), original_size))

    return vocab


def init_vocab(name, data_files, vocab_file, vocab_size, tokenizer, num_workers=1):
    vocab = None
    if vocab_file is not None:
        # If given, load existing word dictionary.
        print('Reading ' + name + ' vocabulary from \'' + vocab_file + '\'...')
        if not opt.load_bpe_voc:
            vocab = onmt.Dict()
        else:
            if name == "target":
                vocab = onmt.Dict([opt.tgt_pad_token, opt.tgt_unk_token,
                                   opt.tgt_bos_token, opt.tgt_eos_token],
                                  lower=opt.lower)
            elif name == "source":
                vocab = onmt.Dict([opt.src_pad_token, opt.src_unk_token,
                                   opt.src_bos_token, opt.src_eos_token],
                                  lower=opt.lower)
            else:
                print("Warning: name should be source or target")
                exit(-1)

        vocab.loadFile(vocab_file)
        print('Loaded ' + str(vocab.size()) + ' ' + name + ' words')

    if vocab is None:
        print('Building ' + name + ' vocabulary...')
        gen_word_vocab = make_vocab(name, data_files, vocab_size, tokenizer, num_workers=num_workers, )

        vocab = gen_word_vocab

    print()
    return vocab


def save_vocabulary(name, vocab, file):
    print('Saving ' + name + ' vocabulary to \'' + file + '\'...')
    vocab.writeFile(file)


def save_dataset(path, data, format, dicts, src_type):
    # Each dataset is comprised of the following components:
    # src: tensors for the source vectors, or the scp_path (in ASR case)
    # tgt: tensors for the target vectors
    # src_lang: tensors for the source language ids (simplified)
    # tgt_lang: tensors for the target language ids (simplified)

    # convert all datasets to pytorch tensors and save to .pt
    if format in ['raw', 'bin']:
        print('Saving data to ' + os.path.join(path, 'data.pt') + '...')

        save_data = {'type': opt.src_type   ,
                     'data': data}
        torch.save(save_data, os.path.join(path, 'data.pt'))
        print("Done")

    # for ASR only
    elif format in ['scp', 'scpmem', 'wav']:
        print('Saving target data to memory indexed data files. Source data is stored only as scp path.')
        from onmt.data.mmap_indexed_dataset import MMapIndexedDatasetBuilder

        assert opt.asr, "ASR data format is required for this memory indexed format"

        # TODO: changing this to before saving everything
        # torch.save(dicts, opt.save_data + '.dict.pt')

        # binarize the training set first
        for set_ in ['tgt', 'src_lang', 'tgt_lang', 'src_atb', 'tgt_atb']:
            if set_ not in data or data[set_] is None:
                continue

            if opt.data_type == 'int64':
                dtype = np.int64
            else:
                dtype = np.int32

            indexed_data = MMapIndexedDatasetBuilder(os.path.join(path, "data.%s.bin" % set_), dtype=dtype)

            # add item from training data to the indexed data
            for tensor in data[set_]:
                indexed_data.add_item(tensor)

            indexed_data.finalize(os.path.join(path, "data.%s.idx" % set_))

            del indexed_data

        for set_ in ['src_sizes', 'tgt_sizes']:

            if data[set_] is not None:

                np_array = np.asarray(data[set_])
                np.save(os.path.join(path, "data.%s.npy") % set_, np_array)
            else:
                print("Training %s not found " % set_)

        # Finally save the audio path
        torch.save(data['src'], os.path.join(path, 'data.scp_path.pt'))
        if 'prev_src' in data and data['prev_src'] is not None:
            torch.save(data['prev_src'], os.path.join(path, 'data.prev_scp_path.pt'))

        print("Done")

    elif opt.format in ['mmap', 'mmem']:
        print('Saving data to memory indexed data files')
        from onmt.data.mmap_indexed_dataset import MMapIndexedDatasetBuilder

        if opt.asr:
            print("ASR data format isn't compatible with memory indexed format")
            raise AssertionError

        # save dicts in this format
        # torch.save(dicts, opt.save_data + '.dict.pt')

        # binarize the training set first
        for set_ in ['src', 'tgt', 'src_lang', 'tgt_lang', 'src_atb', 'tgt_atb']:
            if set_ not in data or data[set_] is None:
                continue

            if opt.data_type == 'int64':
                dtype = np.int64
            else:
                dtype = np.int32

            indexed_data = MMapIndexedDatasetBuilder(os.path.join(path, "data.%s.bin" % set_), dtype=dtype)

            # add item from training data to the indexed data
            for tensor in data[set_]:
                indexed_data.add_item(tensor)

            indexed_data.finalize(os.path.join(path, "data.%s.idx" % set_))

            del indexed_data

        for set_ in ['src_sizes', 'tgt_sizes']:

            if data[set_] is not None:

                np_array = np.asarray(data[set_])
                np.save(os.path.join(path, "data.%s.npy" % set_), np_array)
            else:
                print("Set %s not found " % set_)


def make_lm_data(tgt_file, tgt_dicts, max_tgt_length=1000, input_type='word', data_type='int32'):
    tgt = []
    sizes = []
    count, ignored = 0, 0

    print('Processing %s ...' % (tgt_file))
    tgtf = open(tgt_file)

    eos = torch.LongTensor(1).fill_(opt.tgt_eos_token)
    # print(eos.size())
    tensors = [eos]

    # find the number of words in the sentence
    while True:
        tline = tgtf.readline()

        # normal end of file
        if tline == "":
            break
        tline = tline.strip()
        # source and/or target are empty
        if tline == "":
            print('WARNING: ignoring an empty line (' + str(count + 1) + ')')
            continue

        if input_type == 'word':
            tgt_words = tline.split()
        elif input_type == 'char':
            tgt_words = split_line_by_char(tline)

        tensor = tgt_dicts.convertToIdx(tgt_words,
                                        opt.tgt_unk_token,
                                        None,
                                        opt.tgt_eos_token,
                                        type=data_type)
        # print(tensor.size())
        tensors.append(tensor)

        count = count + 1

        if count % opt.report_every == 0:
            print('... %d sentences prepared' % count)

    tgtf.close()

    # concatenate all tensors into one
    tensor = torch.cat(tensors, dim=-1)

    return tensor


def make_translation_data(src_file, tgt_file, src_dicts, tgt_dicts, tokenizer, max_src_length=64, max_tgt_length=64,
                          add_bos=True, data_type='int64', num_workers=1, verbose=False,
                          external_tokenizer=None, src_lang=None, tgt_lang=None, lang_list=[],
                          early_save=False, savedir="", mirror=False, mirror_savedir=""):
    src, tgt = [], []
    src_sizes = []
    tgt_sizes = []

    if type(lang_list) is dict:
        lang_list = sorted(list(lang_list.keys()))

    print("[INFO] Binarizing file %s ..." % src_file)
    binarized_src = Binarizer.binarize_file(src_file, src_dicts, tokenizer,
                                            bos_word=None, eos_word=None,
                                            data_type=data_type,
                                            num_workers=num_workers, verbose=verbose,
                                            external_tokenizer=external_tokenizer,
                                            lang=src_lang, lang_list=lang_list, target=False
                                            )

    if early_save:
        os.makedirs(savedir, exist_ok=True)
        if mirror:
            os.makedirs(mirror_savedir, exist_ok=True)
        src_len = len(binarized_src['data'])
        print("Saving source data to %s .... with %d entries" % (savedir, src_len))
        if data_type == 'int64':
            dtype = np.int64
        else:
            dtype = np.int32



        from onmt.data.mmap_indexed_dataset import MMapIndexedDatasetBuilder
        indexed_data = MMapIndexedDatasetBuilder(os.path.join(savedir, "data.%s.bin" % "src"), dtype=dtype)

        # add item from training data to the indexed data
        for tensor in binarized_src['data']:
            indexed_data.add_item(tensor)

        indexed_data.finalize(os.path.join(savedir, "data.%s.idx" % "src"))

        del binarized_src['data']
        gc.collect()

        np_array = np.asarray(binarized_src['sizes'])
        np.save(os.path.join(savedir, "data.%s.npy" % "src_sizes"), np_array)

        del binarized_src
        del indexed_data
        del np_array
        gc.collect()

        if mirror:
            print("Saving mirrrored target data to %s .... with %d entries" % (mirror_savedir, src_len))

            source = os.path.join(savedir, "data.%s.bin" % "src")
            target = os.path.join(mirror_savedir, "data.%s.bin" % "tgt")
            os.symlink(os.path.abspath(source), target)

            source = os.path.join(savedir, "data.%s.idx" % "src")
            target = os.path.join(mirror_savedir, "data.%s.idx" % "tgt")
            os.symlink(os.path.abspath(source), target)

            source = os.path.join(savedir, "data.%s.npy" % "src_sizes")
            target = os.path.join(mirror_savedir, "data.%s.npy" % "tgt_sizes")
            os.symlink(os.path.abspath(source), target)

    if add_bos:
        tgt_bos_word = opt.tgt_bos_token
    else:
        tgt_bos_word = None

    print("[INFO] Binarizing file %s ..." % tgt_file)
    binarized_tgt = Binarizer.binarize_file(tgt_file, tgt_dicts, tokenizer,
                                            bos_word=tgt_bos_word, eos_word=opt.tgt_eos_token,
                                            data_type=data_type,
                                            num_workers=num_workers, verbose=verbose,
                                            external_tokenizer=external_tokenizer,
                                            lang=tgt_lang, lang_list=lang_list, target=True
                                            )

    if early_save:

        tgt_len = len(binarized_tgt['data'])
        assert tgt_len == src_len, "Number of samples doesn't match between source and target!!!"

        print("Saving target data to %s .... with %d samples" % (savedir, tgt_len))

        if data_type == 'int64':
            dtype = np.int64
        else:
            dtype = np.int32

        from onmt.data.mmap_indexed_dataset import MMapIndexedDatasetBuilder
        indexed_data = MMapIndexedDatasetBuilder(os.path.join(savedir, "data.%s.bin" % "tgt"), dtype=dtype)

        # add item from training data to the indexed data
        for tensor in binarized_tgt['data']:
            indexed_data.add_item(tensor)

        indexed_data.finalize(os.path.join(savedir, "data.%s.idx" % "tgt"))

        del binarized_tgt['data']
        del indexed_data
        gc.collect()

        np_array = np.asarray(binarized_tgt['sizes'])
        np.save(os.path.join(savedir, "data.%s.npy" % "tgt_sizes"), np_array)

        del binarized_tgt
        del np_array
        gc.collect()

        if mirror:
            print("Saving mirrored source data to %s .... with %d entries" % (mirror_savedir, src_len))

            source = os.path.join(savedir, "data.%s.bin" % "tgt")
            target = os.path.join(mirror_savedir, "data.%s.bin" % "src")
            os.symlink(os.path.abspath(source), target)

            source = os.path.join(savedir, "data.%s.idx" % "tgt")
            target = os.path.join(mirror_savedir, "data.%s.idx" % "src")
            os.symlink(os.path.abspath(source), target)

            source = os.path.join(savedir, "data.%s.npy" % "tgt_sizes")
            target = os.path.join(mirror_savedir, "data.%s.npy" % "src_sizes")
            os.symlink(os.path.abspath(source), target)

        src, tgt, src_sizes, tgt_sizes = None, None, None, None
    else:
        src = binarized_src['data']
        src_sizes = binarized_src['sizes']

        tgt = binarized_tgt['data']
        tgt_sizes = binarized_tgt['sizes']

        # currently we don't ignore anything :D
        ignored = 0

        print(('Prepared %d sentences ' +
               '(%d ignored due to length == 0 or src len > %d or tgt len > %d)') %
              (len(src), ignored, max_src_length, max_tgt_length))

    return src, tgt, src_sizes, tgt_sizes


def make_asr_data(src_file, tgt_file, tgt_dicts, tokenizer,
                  max_src_length=64, max_tgt_length=64, add_bos=True, data_type='int64', num_mel_bin=0,
                  num_workers=1, verbose=False,
                  input_type='word', stride=1, concat=4, prev_context=0, fp16=False, reshape=True,
                  format="wav",
                  external_tokenizer=None, src_lang=None, tgt_lang=None, lang_list=[]):
    src, tgt = [], []
    src_sizes = []
    tgt_sizes = []
    count, ignored = 0, 0
    n_unk_words = 0

    if add_bos:
        tgt_bos_word = opt.tgt_bos_token
    else:
        tgt_bos_word = None

    if tgt_file is not None:
        print("[INFO] Binarizing file %s ..." % tgt_file)

        binarized_tgt = Binarizer.binarize_file(tgt_file, tgt_dicts, tokenizer,
                                                bos_word=tgt_bos_word, eos_word=opt.tgt_eos_token,
                                                data_type=data_type,
                                                num_workers=num_workers, verbose=verbose,
                                                external_tokenizer=external_tokenizer,
                                                lang=tgt_lang, lang_list=lang_list, target=True)

        tgt = binarized_tgt['data']
        tgt_sizes = binarized_tgt['sizes']

        ignored = 0

    else:
        tgt = None
        tgt_sizes = None

    print('[INFO] Processing %s  ...' % src_file)

    # speech binarizer has to be 1 thread at the moment
    binarized_src = SpeechBinarizer.binarize_file(src_file, format=format, concat=concat,
                                                  stride=stride, fp16=fp16, prev_context=prev_context,
                                                  num_workers=num_workers, verbose=verbose, num_mel_bin=num_mel_bin)

    src = binarized_src['data']
    src_sizes = binarized_src['sizes']

    if len(src_sizes) != len(tgt_sizes) and tgt_file is not None:
        print("Warning: data size mismatched. Src: %d . Tgt: %d" % len(src_sizes), len(tgt_sizes))

    print(('Prepared %d sentences ' +
           '(%d ignored due to length == 0 or src len > %d or tgt len > %d)') %
          (len(src), ignored, max_src_length, max_tgt_length))

    return src, tgt, src_sizes, tgt_sizes


def main():
    dicts = {}

    tokenizer = onmt.Tokenizer(opt.input_type, opt.lower)

    # We can load the dictionary from another project to ensure consistency
    if opt.load_dict is not None and len(opt.load_dict) > 0:
        print("[INFO] Loading dictionary from ... %s" % opt.load_dict)
        dicts = torch.load(opt.load_dict)

    # construct set of languages from the training languages
    src_langs = opt.train_src_lang.split("|")
    tgt_langs = opt.train_tgt_lang.split("|")
    langs = (src_langs + tgt_langs)
    langs = sorted(list(set(langs)))

    if len (opt.train_src_atbs) > 0:
        src_atbs = opt.train_src_atbs.split("|")
        tgt_atbs = opt.train_tgt_atbs.split("|")
        atbs = (src_atbs + tgt_atbs)
        atbs = sorted(list(set(atbs)))
    else:
        atbs = []

    if not opt.load_dict:
        dicts['langs'] = dict()

        for lang in langs:
            idx = len(dicts['langs'])
            dicts['langs'][lang] = idx

        dicts['atbs'] = dict()

        for atb in atbs:
            idx = len(dicts['atbs'])
            dicts['atbs'][atb] = idx
    else:

        if 'langs' not in dicts:
            dicts['langs'] = dict()
        else:
            print(dicts['langs'])
            print("Adding languages to existing dictionary ...")

        for lang in langs:
            idx = len(dicts['langs'])
            if lang not in dicts['langs']:
                dicts['langs'][lang] = idx

        if 'atbs' not in dicts:
            dicts['atbs'] = dict()
        else:
            print("Adding attributes to existing dictionary ...")

        for atb in atbs:
            idx = len(dicts['atbs'])
            if atb not in dicts['atbs']:
                dicts['atbs'][atb] = idx

    print("Languages: ", dicts['langs'])
    print("Attributes: ", dicts['atbs'])


    start = time.time()

    src_train_files = opt.train_src.split("|")
    tgt_train_files = opt.train_tgt.split("|")
    # for ASR and LM we only need to build vocab for the 'target' language
    if opt.asr or opt.lm:
        dicts['tgt'] = init_vocab('target', tgt_train_files, opt.tgt_vocab,
                                  opt.tgt_vocab_size, tokenizer, num_workers=opt.num_threads)
    elif opt.join_vocab:
        dicts['src'] = init_vocab('source', set(src_train_files + tgt_train_files), opt.src_vocab,
                                  opt.tgt_vocab_size, tokenizer, num_workers=opt.num_threads)
        dicts['tgt'] = dicts['src']

    else:
        dicts['src'] = init_vocab('source', src_train_files, opt.src_vocab,
                                  opt.src_vocab_size, tokenizer, num_workers=opt.num_threads)

        dicts['tgt'] = init_vocab('target', tgt_train_files, opt.tgt_vocab,
                                  opt.tgt_vocab_size, tokenizer, num_workers=opt.num_threads)

    elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
    print("Vocabulary generated after %s" % elapse)

    if opt.lm:
        print('Preparing training language model ...')
        train = dict()
        train['tgt'] = make_lm_data(opt.train_tgt,
                                    dicts['tgt'])
        train['src'] = None

        valid = dict()
        valid['tgt'] = make_lm_data(opt.valid_tgt,
                                    dicts['tgt'])
        valid['src'] = None
        train['src_sizes'] = None
        train['tgt_sizes'] = None
        valid['src_sizes'] = None
        valid['tgt_sizes'] = None

    elif opt.asr:
        print('Preparing training acoustic model ...')

        src_input_files = opt.train_src.split("|")
        tgt_input_files = opt.train_tgt.split("|")

        src_langs = opt.train_src_lang.split("|")
        tgt_langs = opt.train_tgt_lang.split("|")
        src_atbs = opt.train_src_atbs.split("|") if len(atbs) > 0 else [None] * len(src_input_files)
        tgt_atbs = opt.train_tgt_atbs.split("|") if len(atbs) > 0 else [None] * len(tgt_input_files)

        assert len(src_input_files) == len(src_langs)
        assert len(src_input_files) == len(src_atbs)
        assert len(src_input_files) == len(tgt_input_files)
        assert len(tgt_input_files) == len(tgt_langs)
        assert len(tgt_input_files) == len(tgt_atbs)

        past_src_files = opt.past_train_src.split("|")
        idx = 0
        n_input_files = len(src_input_files)

        # Training data  ###################################################################

        train = dict()
        train['src'], train['tgt'] = list(), list()
        train['src_sizes'], train['tgt_sizes'] = list(), list()
        train['src_atb'], train['tgt_atb'] = list(), list()
        train['src_lang'], train['tgt_lang'] = list(), list()

        data = dict()

        if opt.past_train_src and len(past_src_files) == len(src_input_files):
            train['past_src'] = list()
            train['past_src_sizes'] = list()

        for i, (src_file, tgt_file, src_lang, tgt_lang, src_atb, tgt_atb) in \
                enumerate(zip(src_input_files, tgt_input_files, src_langs, tgt_langs, src_atbs, tgt_atbs)):

            data_name = "train.%i.%s-%s" % (idx, src_lang, tgt_lang)
            dataset_path = os.path.join(dirname(opt.save_data), data_name)
            if opt.multi_dataset and opt.resume:
                print("Checking existing path %s ..." % dataset_path)
                if os.path.exists(dataset_path):
                    print("[INFO] Found data %s in the savedir ... Ignoring" % data_name)
                    idx = idx + 1
                    continue

            src_data, tgt_data, src_sizes, tgt_sizes = make_asr_data(src_file, tgt_file,
                                                                     dicts['tgt'], tokenizer,
                                                                     max_src_length=opt.src_seq_length,
                                                                     max_tgt_length=opt.tgt_seq_length,
                                                                     input_type=opt.input_type,
                                                                     stride=opt.stride, concat=opt.concat,
                                                                     prev_context=opt.previous_context,
                                                                     fp16=opt.fp16,
                                                                     add_bos=not opt.no_bos,
                                                                     format=opt.asr_format,
                                                                     num_mel_bin=opt.num_mel_bin,
                                                                     num_workers=opt.num_threads,
                                                                     external_tokenizer=opt.external_tokenizer,
                                                                     tgt_lang=tgt_lang, verbose=opt.verbose,
                                                                     lang_list=dicts['langs'],)

            n_samples = len(src_data)
            src_atb_data, tgt_atb_data = None, None
            if n_input_files == 1 or opt.multi_dataset:
                # For single-file cases we only need to have 1 language per file
                # which will be broadcasted
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]])]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]])]

                # by default its 0
                if len(atbs) > 0:
                    src_atb_data = [torch.Tensor([dicts['atbs'][src_atb]])]
                    tgt_atb_data = [torch.Tensor([dicts['atbs'][tgt_atb]])]

            else:
                # each sample will have a different language id
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]]) for _ in range(n_samples)]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]]) for _ in range(n_samples)]

                if len(atbs) > 0:
                    src_atb_data = [torch.Tensor([dicts['atbs'][src_atb]]) for _ in range(n_samples)]
                    tgt_atb_data = [torch.Tensor([dicts['atbs'][tgt_atb]]) for _ in range(n_samples)]

            # processing the previous segment
            if opt.past_train_src and len(past_src_files) == len(src_input_files):
                past_src_file = past_src_files[i]

                past_src_data, _, past_src_sizes, _ = make_asr_data(past_src_file, None, None, None,
                                                                    input_type=opt.input_type,
                                                                    stride=opt.stride, concat=opt.concat,
                                                                    prev_context=opt.previous_context,
                                                                    add_bos=not opt.no_bos,
                                                                    fp16=opt.fp16,
                                                                    num_mel_bin=opt.num_mel_bin,
                                                                    format=opt.asr_format,
                                                                    num_workers=opt.num_threads,
                                                                    external_tokenizer=opt.external_tokenizer,
                                                                    tgt_lang=tgt_lang, verbose=opt.verbose,
                                                                    lang_list=dicts['langs'])

                if opt.multi_dataset:
                    data['prev_src'] = prev_src_data
                else:
                    train['past_src'] += past_src_data
                    train['past_src_sizes'] += past_src_sizes

            # Finalizing Training data  ###################################################################

            if opt.multi_dataset:

                data['src'] = src_data
                data['tgt'] = tgt_data

                data['src_sizes'] = src_sizes
                data['tgt_sizes'] = tgt_sizes
                data['src_lang'] = src_lang_data
                data['tgt_lang'] = tgt_lang_data

                if len(atbs) > 0:
                    data['src_atb'] = src_atb_data
                    data['tgt_atb'] = tgt_atb_data
                print("Saving training set %i %s-%s to disk ..." % (idx, src_lang, tgt_lang))

                # take basedir from opt.save_data
                path = os.path.join(dirname(opt.save_data), "train.%i.%s-%s" % (idx, src_lang, tgt_lang))
                os.makedirs(path, exist_ok=True)

                # save data immediately
                # TODO: save the prev src as well
                save_dataset(path, data, opt.format, dicts, opt.src_type)
                idx = idx + 1

                del data
                data = dict()

            else:
                train['src'] += src_data
                train['tgt'] += tgt_data
                train['src_sizes'] += src_sizes
                train['tgt_sizes'] += tgt_sizes
                train['src_lang'] += src_lang_data
                train['tgt_lang'] += tgt_lang_data
                if len(atbs) > 0:
                    train['src_atb'] += src_atb_data
                    train['tgt_atb'] += tgt_atb_data

        # Validation data  ###################################################################

        print('Preparing validation ...')

        src_input_files = opt.valid_src.split("|")
        tgt_input_files = opt.valid_tgt.split("|")
        past_src_files = opt.past_valid_src.split("|")

        src_langs = opt.valid_src_lang.split("|")
        tgt_langs = opt.valid_tgt_lang.split("|")
        src_atbs = opt.valid_src_atbs.split("|") if len(atbs) > 0 else [None] * len(src_input_files)
        tgt_atbs = opt.valid_tgt_atbs.split("|") if len(atbs) > 0 else [None] * len(tgt_input_files)

        assert len(src_input_files) == len(src_langs)
        assert len(src_input_files) == len(tgt_input_files)
        assert len(tgt_input_files) == len(tgt_langs)
        idx = 0
        n_input_files = len(src_input_files)

        data = dict()
        valid = dict()
        valid['src'], valid['tgt'] = list(), list()
        valid['src_sizes'], valid['tgt_sizes'] = list(), list()
        valid['src_lang'], valid['tgt_lang'] = list(), list()
        valid['src_atb'], valid['tgt_atb'] = list(), list()

        if opt.past_train_src and len(past_src_files) == len(src_input_files):
            valid['past_src'] = list()
            valid['past_src_sizes'] = list()

        for i, (src_file, tgt_file, src_lang, tgt_lang, src_atb, tgt_atb) in \
                enumerate(zip(src_input_files, tgt_input_files, src_langs, tgt_langs, src_atbs, tgt_atbs)):

            data_name = "valid.%i.%s-%s" % (idx, src_lang, tgt_lang)
            dataset_path = os.path.join(dirname(opt.save_data), data_name)
            if opt.multi_dataset and opt.resume:
                if os.path.exists(dataset_path):
                    print("[INFO] Found data %s in the savedir ... Ignoring" % data_name)
                    idx = idx + 1
                    continue

            src_data, tgt_data, src_sizes, tgt_sizes = make_asr_data(src_file, tgt_file,
                                                                     dicts['tgt'], tokenizer,
                                                                     max_src_length=max(1024, opt.src_seq_length),
                                                                     max_tgt_length=max(1024, opt.tgt_seq_length),
                                                                     input_type=opt.input_type,
                                                                     stride=opt.stride, concat=opt.concat,
                                                                     prev_context=opt.previous_context,
                                                                     fp16=opt.fp16,
                                                                     num_mel_bin=opt.num_mel_bin,
                                                                     add_bos=not opt.no_bos,
                                                                     format=opt.asr_format,
                                                                     external_tokenizer=opt.external_tokenizer,
                                                                     tgt_lang=tgt_lang, verbose=opt.verbose,
                                                                     lang_list=dicts['langs'],
                                                                     num_workers=opt.num_threads,
                                                                     )

            n_samples = len(src_data)
            if n_input_files == 1 or opt.multi_dataset:
                # For single-file cases we only need to have 1 language per file
                # which will be broadcasted
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]])]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]])]

                # by default its 0
                if len(atbs) > 0:
                    src_atb_data = [torch.Tensor([dicts['atbs'][src_atb]])]
                    tgt_atb_data = [torch.Tensor([dicts['atbs'][tgt_atb]])]
            else:
                # each sample will have a different language id
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]]) for _ in range(n_samples)]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]]) for _ in range(n_samples)]

                if len(atbs) > 0:
                    src_atb_data = [torch.Tensor([dicts['atbs'][src_atb]]) for _ in range(n_samples)]
                    tgt_atb_data = [torch.Tensor([dicts['atbs'][tgt_atb]]) for _ in range(n_samples)]

            # validation past file
            if opt.past_train_src and len(past_src_files) == len(src_input_files):
                past_src_file = past_src_files[i]

                past_src_data, _, past_src_sizes, _ = make_asr_data(past_src_file, None, None, None,
                                                                    input_type=opt.input_type,
                                                                    stride=opt.stride, concat=opt.concat,
                                                                    prev_context=opt.previous_context,
                                                                    fp16=opt.fp16,
                                                                    add_bos=not opt.no_bos,
                                                                    num_mel_bin=opt.num_mel_bin,
                                                                    format=opt.asr_format,
                                                                    num_workers=opt.num_threads,
                                                                    external_tokenizer=opt.external_tokenizer,
                                                                    tgt_lang=tgt_lang, verbose=opt.verbose,
                                                                    lang_list=dicts['langs'])

                valid['past_src'] += past_src_data
                valid['past_src_sizes'] += past_src_sizes

            # Finalizing Validation data ... #########################

            if opt.multi_dataset:
                data['src'] = src_data
                data['tgt'] = tgt_data

                data['src_sizes'] = src_sizes
                data['tgt_sizes'] = tgt_sizes
                data['src_lang'] = src_lang_data
                data['tgt_lang'] = tgt_lang_data
                if len(atbs) > 0:
                    data['src_atb'] = src_atb_data
                    data['tgt_atb'] = tgt_atb_data

                print("Saving validation set %i %s-%s to disk ..." % (idx, src_lang, tgt_lang))

                # take basedir from opt.save_data
                path = os.path.join(dirname(opt.save_data), "valid.%i.%s-%s" % (idx, src_lang, tgt_lang))
                os.makedirs(path, exist_ok=True)

                # save data immediately
                save_dataset(path, data, opt.format, dicts, opt.src_type)
                idx = idx + 1

                del data
                data = dict()
            else:
                valid['src'] += src_data
                valid['tgt'] += tgt_data
                valid['src_sizes'] += src_sizes
                valid['tgt_sizes'] += tgt_sizes
                valid['src_lang'] += src_lang_data
                valid['tgt_lang'] += tgt_lang_data
                if len(atbs) > 0:
                    valid['src_atb'] += src_atb_data
                    valid['tgt_atb'] += tgt_atb_data

    else:   # MACHINE TRANSLATION DATA

        src_input_files = opt.train_src.split("|")
        tgt_input_files = opt.train_tgt.split("|")

        src_langs = opt.train_src_lang.split("|")
        tgt_langs = opt.train_tgt_lang.split("|")

        assert len(src_input_files) == len(src_langs)
        assert len(src_input_files) == len(tgt_input_files)
        assert len(tgt_input_files) == len(tgt_langs)

        past_src_files = opt.past_train_src.split("|")

        n_input_files = len(src_input_files)

        idx = 0
        data = dict()
        train = dict()
        train['src'], train['tgt'] = list(), list()
        train['src_sizes'], train['tgt_sizes'] = list(), list()
        train['src_lang'], train['tgt_lang'] = list(), list()

        if opt.past_train_src and len(past_src_files) == len(src_input_files):
            train['past_src'] = list()
            train['past_src_sizes'] = list()

        start = time.time()
        print('Binarizing data to train translation models...')

        for i, (src_file, tgt_file, src_lang, tgt_lang) in \
                enumerate(zip(src_input_files, tgt_input_files, src_langs, tgt_langs)):

            dataset_idx = idx if not opt.multi_mirror else 2 * idx
            data_name = "train.%i.%s-%s" % (dataset_idx , src_lang, tgt_lang)
            mirrored_data_name = "train.%i.%s-%s" % (dataset_idx + 1 , tgt_lang, src_lang)
            dataset_path = os.path.join(dirname(opt.save_data), data_name)
            mirrored_dataset_path = os.path.join(dirname(opt.save_data), mirrored_data_name)
            if opt.multi_dataset and opt.resume:
                print("Checking existing path %s ..." % dataset_path)
                if os.path.exists(dataset_path):
                    print("[INFO] Found data %s in the savedir ... Ignoring" % data_name)
                    idx = idx + 1
                    continue
            else:
                os.makedirs(dataset_path, exist_ok=True)

            src_data, tgt_data, src_sizes, tgt_sizes = make_translation_data(src_file, tgt_file,
                                                                             dicts['src'], dicts['tgt'], tokenizer,
                                                                             max_src_length=opt.src_seq_length,
                                                                             max_tgt_length=opt.tgt_seq_length,
                                                                             add_bos=(not opt.no_bos),
                                                                             data_type=opt.data_type,
                                                                             num_workers=opt.num_threads,
                                                                             verbose=opt.verbose,
                                                                             external_tokenizer=opt.external_tokenizer,
                                                                             src_lang=src_lang,
                                                                             tgt_lang=tgt_lang,
                                                                             lang_list=dicts['langs'],
                                                                             early_save=opt.multi_dataset,
                                                                             savedir=dataset_path,
                                                                             mirror=opt.multi_mirror,
                                                                             mirror_savedir=mirrored_dataset_path)

            #TODO: check
            # if n_input_files == 1:
            if n_input_files == 1 or opt.multi_dataset:
                # For single-file cases we only need to have 1 language per file
                # which will be broadcasted
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]])]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]])]
            else:
                assert src_data is not None
                n_samples = len(src_data)

                # each sample will have a different language id
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]]) for _ in range(n_samples)]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]]) for _ in range(n_samples)]

            # processing the previous segment
            if opt.past_train_src and len(past_src_files) == len(src_input_files):
                past_src_file = past_src_files[i]

                past_src_data, _, past_src_sizes, _ = make_translation_data(past_src_file, '/dev/null',
                                                                            dicts['src'], dicts['src'], tokenizer,
                                                                            max_src_length=opt.src_seq_length,
                                                                            max_tgt_length=opt.tgt_seq_length,
                                                                            add_bos=(not opt.no_bos),
                                                                            data_type=opt.data_type,
                                                                            num_workers=opt.num_threads,
                                                                            verbose=opt.verbose,
                                                                            external_tokenizer=opt.external_tokenizer,
                                                                            src_lang=src_lang,
                                                                            tgt_lang=tgt_lang,
                                                                            lang_list=dicts['langs'])

                if opt.multi_dataset:
                    data['prev_src'] = prev_src_data
                else:
                    train['past_src'] += past_src_data
                    train['past_src_sizes'] += past_src_sizes

            if opt.multi_dataset:

                data['src'] = src_data
                data['tgt'] = tgt_data

                data['src_sizes'] = src_sizes
                data['tgt_sizes'] = tgt_sizes
                data['src_lang'] = src_lang_data
                data['tgt_lang'] = tgt_lang_data
                print("Saving training set %i %s-%s to disk ..." % (dataset_idx, src_lang, tgt_lang))

                # take basedir from opt.save_data
                path = dataset_path
                os.makedirs(path, exist_ok=True)

                # save data immediately
                # TODO: save the prev src as well
                save_dataset(path, data, opt.format, dicts, opt.src_type)

                if opt.multi_mirror:
                    mdata = dict()
                    mdata['src'] = tgt_data
                    mdata['tgt'] = src_data

                    mdata['tgt_sizes'] = src_sizes
                    mdata['src_sizes'] = tgt_sizes
                    mdata['tgt_lang'] = src_lang_data
                    mdata['src_lang'] = tgt_lang_data
                    print("Saving training set %i %s-%s to disk ..." % (dataset_idx + 1, tgt_lang, src_lang))

                    # take basedir from opt.save_data
                    path = mirrored_dataset_path
                    os.makedirs(path, exist_ok=True)

                    # save data immediately
                    # TODO: save the prev src as well
                    save_dataset(path, mdata, opt.format, dicts, opt.src_type)

                idx = idx + 1

                del data
                data = dict()

            else:
                train['src'] += src_data
                train['tgt'] += tgt_data
                train['src_sizes'] += src_sizes
                train['tgt_sizes'] += tgt_sizes
                train['src_lang'] += src_lang_data
                train['tgt_lang'] += tgt_lang_data

        print('Preparing validation ...')

        src_input_files = opt.valid_src.split("|")
        tgt_input_files = opt.valid_tgt.split("|")
        past_src_files = opt.past_valid_src.split("|")

        src_langs = opt.valid_src_lang.split("|")
        tgt_langs = opt.valid_tgt_lang.split("|")

        assert len(src_input_files) == len(src_langs)
        assert len(src_input_files) == len(tgt_input_files)
        assert len(tgt_input_files) == len(tgt_langs)

        n_input_files = len(src_input_files)

        idx = 0
        data = dict()
        valid = dict()
        valid['src'], valid['tgt'] = list(), list()
        valid['src_sizes'], valid['tgt_sizes'] = list(), list()
        valid['src_lang'], valid['tgt_lang'] = list(), list()

        if opt.past_train_src and len(past_src_files) == len(src_input_files):
            valid['past_src'] = list()
            valid['past_src_sizes'] = list()

        for (src_file, tgt_file, src_lang, tgt_lang) in zip(src_input_files, tgt_input_files, src_langs, tgt_langs):

            src_data, tgt_data, src_sizes, tgt_sizes = make_translation_data(src_file, tgt_file,
                                                                             dicts['src'], dicts['tgt'], tokenizer,
                                                                             max_src_length=max(1024,
                                                                                                opt.src_seq_length),
                                                                             max_tgt_length=max(1024,
                                                                                                opt.tgt_seq_length),
                                                                             add_bos=(not opt.no_bos),
                                                                             data_type=opt.data_type,
                                                                             num_workers=opt.num_threads,
                                                                             verbose=opt.verbose,
                                                                             external_tokenizer=opt.external_tokenizer,
                                                                             src_lang=src_lang,
                                                                             tgt_lang=tgt_lang,
                                                                             lang_list=dicts['langs']
                                                                             )

            n_samples = len(src_data)
            #TODO: this has to be changed
            # if n_input_files == 1:
            if n_input_files == 1 or opt.multi_dataset:
                # For single-file cases we only need to have 1 language per file
                # which will be broadcasted
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]])]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]])]
            else:
                # each sample will have a different language id
                src_lang_data = [torch.Tensor([dicts['langs'][src_lang]]) for _ in range(n_samples)]
                tgt_lang_data = [torch.Tensor([dicts['langs'][tgt_lang]]) for _ in range(n_samples)]

            # validation past file
            if opt.past_train_src and len(past_src_files) == len(src_input_files):
                past_src_file = past_src_files[i]

                past_src_data, _, past_src_sizes, _ = make_translation_data(past_src_file, '/dev/null',
                                                                            dicts['src'], dicts['src'], tokenizer,
                                                                            max_src_length=max(1024,
                                                                                               opt.src_seq_length),
                                                                            max_tgt_length=max(1024,
                                                                                               opt.tgt_seq_length),
                                                                            add_bos=(not opt.no_bos),
                                                                            data_type=opt.data_type,
                                                                            num_workers=opt.num_threads,
                                                                            verbose=opt.verbose,
                                                                            external_tokenizer=opt.external_tokenizer,
                                                                            src_lang=src_lang,
                                                                            tgt_lang=tgt_lang,
                                                                            lang_list=dicts['langs'])

                valid['past_src'] += past_src_data
                valid['past_src_sizes'] += past_src_sizes

            if opt.multi_dataset:
                data['src'] = src_data
                data['tgt'] = tgt_data

                data['src_sizes'] = src_sizes
                data['tgt_sizes'] = tgt_sizes
                data['src_lang'] = src_lang_data
                data['tgt_lang'] = tgt_lang_data

                print("Saving validation set %i %s-%s to disk ..." % (idx, src_lang, tgt_lang))

                # take basedir from opt.save_data
                path = os.path.join(dirname(opt.save_data), "valid.%i.%s-%s" % (idx, src_lang, tgt_lang))
                os.makedirs(path, exist_ok=True)

                # save data immediately
                save_dataset(path, data, opt.format, dicts, opt.src_type)
                idx = idx + 1
            else:
                valid['src'] += src_data
                valid['tgt'] += tgt_data
                valid['src_sizes'] += src_sizes
                valid['tgt_sizes'] += tgt_sizes
                valid['src_lang'] += src_lang_data
                valid['tgt_lang'] += tgt_lang_data

        elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
        print("Binarization finished after %s" % elapse)

    if opt.src_vocab is None and opt.asr == False and opt.lm == False:
        save_vocabulary('source', dicts['src'], opt.save_data + '.src.dict')
    if opt.tgt_vocab is None:
        save_vocabulary('target', dicts['tgt'], opt.save_data + '.tgt.dict')

    if opt.multi_dataset:
        # SAVE DATA
        print("Saving dictionary to %s" % (opt.save_data + '.dict.pt'))
        torch.save(dicts, opt.save_data + '.dict.pt')

        if opt.src_vocab is None and opt.asr == False and opt.lm == False:
            save_vocabulary('source', dicts['src'], opt.save_data + '.src.dict')
        if opt.tgt_vocab is None:
            save_vocabulary('target', dicts['tgt'], opt.save_data + '.tgt.dict')

        print("Finished.")
    else:
        if opt.format in ['raw', 'bin']:

            print('Saving data to \'' + opt.save_data + '.train.pt\'...')

            save_data = {'dicts': dicts,
                         'type': opt.src_type,
                         'train': train,
                         'valid': valid}
            torch.save(save_data, opt.save_data + '.train.pt')
            print("Done")

        elif opt.format in ['scp', 'scpmem', 'wav']:
            print('Saving target data to memory indexed data files. Source data is stored only as scp path.')
            from onmt.data.mmap_indexed_dataset import MMapIndexedDatasetBuilder

            assert opt.asr, "ASR data format is required for this memory indexed format"

            torch.save(dicts, opt.save_data + '.dict.pt')

            # binarize the training set first
            for set_ in ['tgt', 'src_lang', 'tgt_lang']:
                if train[set_] is None:
                    continue

                if opt.data_type == 'int64':
                    dtype = np.int64
                else:
                    dtype = np.int32

                train_data = MMapIndexedDatasetBuilder(opt.save_data + ".train.%s.bin" % set_, dtype=dtype)

                # add item from training data to the indexed data
                for tensor in train[set_]:
                    train_data.add_item(tensor)

                train_data.finalize(opt.save_data + ".train.%s.idx" % set_)

                del train_data

                if valid[set_] is None:
                    continue

                valid_data = MMapIndexedDatasetBuilder(opt.save_data + ".valid.%s.bin" % set_, dtype=dtype)

                # add item from training data to the indexed data
                for tensor in valid[set_]:
                    valid_data.add_item(tensor)

                valid_data.finalize(opt.save_data + ".valid.%s.idx" % set_)

                del valid_data

            for set_ in ['src_sizes', 'tgt_sizes']:

                if train[set_] is not None:

                    np_array = np.asarray(train[set_])
                    np.save(opt.save_data + ".train.%s.npy" % set_, np_array)
                else:
                    print("Training %s not found " % set_)

                if valid[set_] is not None:

                    np_array = np.asarray(valid[set_])
                    np.save(opt.save_data + ".valid.%s.npy" % set_, np_array)
                else:
                    print("Validation %s not found " % set_)

            if 'past_src' in train and len(train['past_src']) > 0:
                set_ = 'past_src_sizes'

                if train[set_] is not None:

                    np_array = np.asarray(train[set_])
                    np.save(opt.save_data + ".train.%s.npy" % set_, np_array)
                else:
                    print("Training %s not found " % set_)

                if valid[set_] is not None:

                    np_array = np.asarray(valid[set_])
                    np.save(opt.save_data + ".valid.%s.npy" % set_, np_array)
                else:
                    print("Validation %s not found " % set_)

            # Finally save the audio path
            save_data = {'train': train['src'],
                         'valid': valid['src']}

            # remember to take into account the past information
            if 'past_src' in train and len(train['past_src']) > 0:
                save_data['train_past'] = train['past_src']
                save_data['valid_past'] = valid['past_src']

            if opt.format in ['wav']:
                torch.save(save_data, opt.save_data + '.wav_path.pt')
            else:
                torch.save(save_data, opt.save_data + '.scp_path.pt')

            print("Done")

        elif opt.format in ['mmap', 'mmem']:
            print('Saving data to memory indexed data files')
            from onmt.data.mmap_indexed_dataset import MMapIndexedDatasetBuilder


            # save dicts in this format
            torch.save(dicts, opt.save_data + '.dict.pt')

            # binarize the training set first
            for set_ in ['src', 'tgt', 'src_lang', 'tgt_lang', 'past_src']:
                if set_ not in train or train[set_] is None:
                    continue

                if opt.data_type == 'int64':
                    dtype = np.int64
                else:
                    dtype = np.int32

                train_data = MMapIndexedDatasetBuilder(opt.save_data + ".train.%s.bin" % set_, dtype=dtype)

                # add item from training data to the indexed data
                for tensor in train[set_]:
                    train_data.add_item(tensor)

                train_data.finalize(opt.save_data + ".train.%s.idx" % set_)

                del train_data

                if valid[set_] is None:
                    continue

                valid_data = MMapIndexedDatasetBuilder(opt.save_data + ".valid.%s.bin" % set_, dtype=dtype)

                # add item from training data to the indexed data
                for tensor in valid[set_]:
                    valid_data.add_item(tensor)

                valid_data.finalize(opt.save_data + ".valid.%s.idx" % set_)

                del valid_data

            for set_ in ['src_sizes', 'tgt_sizes']:

                if set_ not in train or train[set_] is not None:

                    np_array = np.asarray(train[set_])
                    np.save(opt.save_data + ".train.%s.npy" % set_, np_array)
                else:
                    print("Training %s not found " % set_)

            if 'past_src' in train and len(train['past_src']) > 0:
                set_ = 'past_src_sizes'

                if train[set_] is not None:

                    np_array = np.asarray(train[set_])
                    np.save(opt.save_data + ".train.%s.npy" % set_, np_array)
                else:
                    print("Training %s not found " % set_)

                if valid[set_] is not None:

                    np_array = np.asarray(valid[set_])
                    np.save(opt.save_data + ".valid.%s.npy" % set_, np_array)
                else:
                    print("Validation %s not found " % set_)

        else:
            raise NotImplementedError

if __name__ == "__main__":
    main()


def safe_readline(f):
    pos = f.tell()
    while True:
        try:
            return f.readline()
        except UnicodeDecodeError:
            pos -= 1
            f.seek(pos)  # search where this character begins