from argparse import ArgumentParser
from pathlib import Path
from tqdm import tqdm, trange
from tempfile import TemporaryDirectory
import shelve
from multiprocessing import Pool

from random import random, randrange, randint, shuffle, choice
from pytorch_transformers.tokenization_auto import AutoTokenizer
import numpy as np
import json
import collections
from natsort import natsorted
import glob

class DocumentDatabase:
    def __init__(self, reduce_memory=False):
        if reduce_memory:
            self.temp_dir = TemporaryDirectory()
            self.working_dir = Path(
            self.document_shelf_filepath = self.working_dir / 'shelf.db'
            self.document_shelf =,
                                              flag='n', protocol=-1)
            self.documents = None
            self.documents = []
            self.document_shelf = None
            self.document_shelf_filepath = None
            self.temp_dir = None
        self.doc_lengths = []
        self.doc_cumsum = None
        self.cumsum_max = None
        self.reduce_memory = reduce_memory

    def add_document(self, document):
        if not document:
        if self.reduce_memory:
            current_idx = len(self.doc_lengths)
            self.document_shelf[str(current_idx)] = document

    def _precalculate_doc_weights(self):
        self.doc_cumsum = np.cumsum(self.doc_lengths)
        self.cumsum_max = self.doc_cumsum[-1]

    def sample_doc(self, current_idx, sentence_weighted=True):
        # Uses the current iteration counter to ensure we don't sample the same doc twice
        if sentence_weighted:
            # With sentence weighting, we sample docs proportionally to their sentence length
            if self.doc_cumsum is None or len(self.doc_cumsum) != len(self.doc_lengths):
            rand_start = self.doc_cumsum[current_idx]
            rand_end = rand_start + self.cumsum_max - self.doc_lengths[current_idx]
            sentence_index = randrange(rand_start, rand_end) % self.cumsum_max
            sampled_doc_index = np.searchsorted(self.doc_cumsum, sentence_index, side='right')
            # If we don't use sentence weighting, then every doc has an equal chance to be chosen
            sampled_doc_index = (current_idx + randrange(1, len(self.doc_lengths))) % len(self.doc_lengths)
        assert sampled_doc_index != current_idx
        if self.reduce_memory:
            return self.document_shelf[str(sampled_doc_index)]
            return self.documents[sampled_doc_index]

    def __len__(self):
        return len(self.doc_lengths)

    def __getitem__(self, item):
        if self.reduce_memory:
            return self.document_shelf[str(item)]
            return self.documents[item]

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_val, traceback):
        if self.document_shelf is not None:
        if self.temp_dir is not None:

def truncate_seq_pair(tokens_a, tokens_b, max_num_tokens):
    """Truncates a pair of sequences to a maximum sequence length. Lifted from Google's BERT repo."""
    while True:
        total_length = len(tokens_a) + len(tokens_b)
        if total_length <= max_num_tokens:

        trunc_tokens = tokens_a if len(tokens_a) > len(tokens_b) else tokens_b
        assert len(trunc_tokens) >= 1

        # We want to sometimes truncate from the front and sometimes from the
        # back to add more randomness and avoid biases.
        if random() < 0.5:
            del trunc_tokens[0]

MaskedLmInstance = collections.namedtuple("MaskedLmInstance",
                                          ["index", "label"])

def create_masked_lm_predictions(tokens, masked_lm_prob, max_predictions_per_seq, whole_word_mask, tokenizer):
    """Creates the predictions for the masked LM objective. This is mostly copied from the Google BERT repo, but
    with several refactors to clean it up and remove a lot of unnecessary variables."""
    cand_indices = []
    for (i, token) in enumerate(tokens):
        if token == tokenizer.cls_token or token == tokenizer.sep_token:
        # Whole Word Masking means that if we mask all of the wordpieces
        # corresponding to an original word. When a word has been split into
        # WordPieces, the first token does not have any marker and any subsequence
        # tokens are prefixed with ##. So whenever we see the ## token, we
        # append it to the previous set of word indexes.
        # Note that Whole Word Masking does *not* change the training code
        # at all -- we still predict each WordPiece independently, softmaxed
        # over the entire vocabulary.
        if (whole_word_mask and len(cand_indices) >= 1 and token.startswith("##")):

    num_to_mask = min(max_predictions_per_seq,
                      max(1, int(round(len(tokens) * masked_lm_prob))))
    masked_lms = []
    covered_indexes = set()
    for index_set in cand_indices:
        if len(masked_lms) >= num_to_mask:
        # If adding a whole-word mask would exceed the maximum number of
        # predictions, then just skip this candidate.
        if len(masked_lms) + len(index_set) > num_to_mask:
        is_any_index_covered = False
        for index in index_set:
            if index in covered_indexes:
                is_any_index_covered = True
        if is_any_index_covered:
        for index in index_set:

            masked_token = None
            # 80% of the time, replace with [MASK]
            if random() < 0.8:
                masked_token = "[MASK]"
                # 10% of the time, keep original
                if random() < 0.5:
                    masked_token = tokens[index]
                # 10% of the time, replace with random word
                    masked_token = choice(tokenizer.vocab_list)
            masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
            tokens[index] = masked_token

    assert len(masked_lms) <= num_to_mask
    masked_lms = sorted(masked_lms, key=lambda x: x.index)
    mask_indices = [p.index for p in masked_lms]
    masked_token_labels = [p.label for p in masked_lms]

    return tokens, mask_indices, masked_token_labels

def create_instances_from_document(
        doc_database, doc_idx, max_seq_length, short_seq_prob,
        masked_lm_prob, max_predictions_per_seq, whole_word_mask, tokenizer, next_sent_prediction):
    """This code is mostly a duplicate of the equivalent function from Google BERT's repo.
    However, we make some changes and improvements. Sampling is improved and no longer requires a loop in this function.
    Also, documents are sampled proportionally to the number of sentences they contain, which means each sentence
    (rather than each document) has an equal chance of being sampled as a false example for the NextSentence task."""
    document = doc_database[doc_idx]
    # Account for [CLS], [SEP], [SEP]
    max_num_tokens = max_seq_length - 3

    # We *usually* want to fill up the entire sequence since we are padding
    # to `max_seq_length` anyways, so short sequences are generally wasted
    # computation. However, we *sometimes*
    # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter
    # sequences to minimize the mismatch between pre-training and fine-tuning.
    # The `target_seq_length` is just a rough target however, whereas
    # `max_seq_length` is a hard limit.
    target_seq_length = max_num_tokens
    if random() < short_seq_prob:
        target_seq_length = randint(2, max_num_tokens)

    # We DON'T just concatenate all of the tokens from a document into a long
    # sequence and choose an arbitrary split point because this would make the
    # next sentence prediction task too easy. Instead, we split the input into
    # segments "A" and "B" based on the actual "sentences" provided by the user
    # input.
    instances = []
    current_chunk = []
    current_length = 0
    i = 0
    while i < len(document):
        segment = document[i]
        current_length += len(segment)
        if i == len(document) - 1 or current_length >= target_seq_length:
            if current_chunk:
                # `a_end` is how many segments from `current_chunk` go into the `A`
                # (first) sentence.
                a_end = 1
                if len(current_chunk) >= 2:
                    a_end = randrange(1, len(current_chunk))

                tokens_a = []
                for j in range(a_end):

                tokens_b = []

                # Random next
                if len(current_chunk) == 1 or (next_sent_prediction  and random() < 0.5):
                    is_random_next = True
                    target_b_length = target_seq_length - len(tokens_a)

                    # Sample a random document, with longer docs being sampled more frequently
                    random_document = doc_database.sample_doc(current_idx=doc_idx, sentence_weighted=True)

                    random_start = randrange(0, len(random_document))
                    for j in range(random_start, len(random_document)):
                        if len(tokens_b) >= target_b_length:
                    # We didn't actually use these segments so we "put them back" so
                    # they don't go to waste.
                    num_unused_segments = len(current_chunk) - a_end
                    i -= num_unused_segments
                # Actual next
                    is_random_next = False
                    for j in range(a_end, len(current_chunk)):
                        tokens_a.extend(current_chunk[j])  # put the tokens of the second segment in tokens_a
                truncate_seq_pair(tokens_a, tokens_b, max_num_tokens)
                assert len(tokens_a) >= 1
                if next_sent_prediction:
                    tokens = [tokenizer.cls_token] + tokens_a + [tokenizer.sep_token] + tokens_b + [tokenizer.sep_token]
                    # The segment IDs are 0 for the [CLS] token, the A tokens and the first [SEP]
                    # They are 1 for the B tokens and the final [SEP]
                    segment_ids = [0 for _ in range(len(tokens_a) + 2)] + [1 for _ in range(len(tokens_b) + 1)]
                    tokens = [tokenizer.cls_token] + tokens_a + tokens_b + [tokenizer.sep_token]
                    segment_ids = [0] * len(tokens)

                tokens, masked_lm_positions, masked_lm_labels = create_masked_lm_predictions(
                    tokens, masked_lm_prob, max_predictions_per_seq, whole_word_mask, tokenizer)

                instance = {
                    "tokens": tokens,
                    "segment_ids": segment_ids,
                    "is_random_next": is_random_next,
                    "masked_lm_positions": masked_lm_positions,
                    "masked_lm_labels": masked_lm_labels}
            current_chunk = []
            current_length = 0
        i += 1

    return instances

def create_training_file(docs, tokenizer, args, epoch_num):
    epoch_filename = args.output_dir / "epoch_{}.json".format(epoch_num)
    num_instances = 0
    with'w') as epoch_file:
        for doc_idx in trange(len(docs), desc="Document"):
            doc_instances = create_instances_from_document(
                docs, doc_idx, max_seq_length=args.max_seq_len, short_seq_prob=args.short_seq_prob,
                masked_lm_prob=args.masked_lm_prob, max_predictions_per_seq=args.max_predictions_per_seq,
                whole_word_mask=args.do_whole_word_mask, tokenizer=tokenizer,
            doc_instances = [json.dumps(instance) for instance in doc_instances]
            for instance in doc_instances:
                epoch_file.write(instance + '\n')
                num_instances += 1
    metrics_file = args.output_dir / "epoch_{}_metrics.json".format(epoch_num)
    with'w') as metrics_file:
        metrics = {
            "num_training_examples": num_instances,
            "max_seq_len": args.max_seq_len

def input_file_to_training_data(args, input_file, epoch, tokenizer, num_files):
    with DocumentDatabase(reduce_memory=args.reduce_memory) as docs:
        with open(input_file) as f:
            doc = []
            for line in tqdm(f, desc="Loading Dataset", unit=" lines"):
                line = line.strip()
                if line == "":
                    doc = []
                    tokens = tokenizer.tokenize(line)
            if doc:
                docs.add_document(doc)  # If the last doc didn't end on a newline, make sure it still gets added
        if len(docs) <= 1:
            exit("ERROR: No document breaks were found in the input file! These are necessary to allow the script to "
                    "ensure that random NextSentences are not sampled from the same document. Please add blank lines to "
                    "indicate breaks between documents in your input file. If your dataset does not contain multiple "
                    "documents, blank lines can be inserted at any natural boundary, such as the ends of chapters, "
                    "sections or paragraphs.")

        for i in range(args.epochs_to_generate):
            create_training_file(docs, tokenizer, args, epoch + i * num_files)

def main():
    parser = ArgumentParser()
    parser.add_argument('--train_corpus', type=str, required=True, help="Path to training corpus in glob format")
    parser.add_argument("--output_dir", type=Path, required=True)
    parser.add_argument("--bert_model", type=str, required=True,
                        choices=["bert-base-uncased", "bert-large-uncased", "bert-large-cased", "bert-base-cased",
                                 "bert-base-multilingual-uncased", "bert-base-chinese", "bert-base-multilingual-cased",
                                 "roberta-base", "roberta-large"])
    parser.add_argument("--do_lower_case", action="store_true")
    parser.add_argument("--do_whole_word_mask", action="store_true",
                        help="Whether to use whole word masking rather than per-WordPiece masking.")
    parser.add_argument("--reduce_memory", action="store_true",
                        help="Reduce memory usage for large datasets by keeping data on disc rather than in memory")

    parser.add_argument("--num_workers", type=int, default=1,
                        help="The number of workers to use to write the files")
    parser.add_argument("--epochs_to_generate", type=int, default=3,
                        help="Number of epochs of data to pregenerate")
    parser.add_argument("--max_seq_len", type=int, default=128)
    parser.add_argument("--short_seq_prob", type=float, default=0.1,
                        help="Probability of making a short sentence as a training example")
    parser.add_argument("--masked_lm_prob", type=float, default=0.15,
                        help="Probability of masking each token for the LM task")
    parser.add_argument("--max_predictions_per_seq", type=int, default=20,
                        help="Maximum number of tokens to mask in each sequence")
    parser.add_argument("--do_next_sent_prediction", action="store_true",
                        help="Add the next sentence prediction task (as in BERT) or ignore it (as in RoBERTa)")

    args = parser.parse_args()

    if args.num_workers > 1 and args.reduce_memory:
        raise ValueError("Cannot use multiple workers while reducing memory")


    tokenizer = AutoTokenizer.from_pretrained(args.bert_model)
    tokenizer.vocab_list = list((tokenizer.encoder if hasattr(tokenizer, 'encoder') else tokenizer.vocab).keys())

    files = glob.glob(args.train_corpus)
    if args.num_workers > 1:
        pool = Pool(args.num_workers)
        arguments = [(args, input_file, i, tokenizer, len(files)) for i, input_file in enumerate(files)]
        pool.starmap(input_file_to_training_data, arguments)
        for i, input_file in enumerate(tqdm(files)):
            input_file_to_training_data(args, input_file, i, tokenizer, len(files))

if __name__ == '__main__':