# Step 1: Slurp the dataset up, tokenize each sentence, and store as docs -> sentences -> tokens # Step 2: Walk over the dataset, using the Google BERT logic to concatenate sentences into training examples # Step 3: Write out the examples, possibly as Torch tensors? from argparse import ArgumentParser from pathlib import Path from tqdm import tqdm, trange from random import random, randint, shuffle, choice, sample from pytorch_pretrained_bert.tokenization import BertTokenizer import json class DocumentDatabase: def __init__(self, document_list): self.document_list = document_list self.doc_starts = {} self.weighted_doc_samples = [] i = 0 for doc_idx, doc in enumerate(document_list): self.doc_starts[doc_idx] = i self.weighted_doc_samples.extend([doc_idx] * len(doc)) i += len(doc) 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: num_sentences = len(self.document_list[current_idx]) # This very painful line randomly selects a document, weighted by the number of sentences they contain, # while guaranteeing that it won't return the original document sampled_val = ( (self.doc_starts[current_idx] + num_sentences + randint(0, len(self.weighted_doc_samples) - num_sentences - 1)) % len(self.weighted_doc_samples)) sampled_doc_index = self.weighted_doc_samples[sampled_val] else: # If we don't use sentence weighting, then every doc has an equal chance to be chosen sampled_doc_index = current_idx + randint(1, len(self.document_list)-1) assert sampled_doc_index != current_idx return self.document_list[sampled_doc_index] def __len__(self): return len(self.document_list) def __getitem__(self, item): return self.document_list[item] 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: break 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] else: trunc_tokens.pop() def create_masked_lm_predictions(tokens, masked_lm_prob, max_predictions_per_seq, vocab_list): """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 == "[CLS]" or token == "[SEP]": continue cand_indices.append(i) num_to_mask = min(max_predictions_per_seq, max(1, int(round(len(tokens) * masked_lm_prob)))) shuffle(cand_indices) mask_indices = sorted(sample(cand_indices, num_to_mask)) masked_token_labels = [] for index in mask_indices: # 80% of the time, replace with [MASK] if random() < 0.8: masked_token = "[MASK]" else: # 10% of the time, keep original if random() < 0.5: masked_token = tokens[index] # 10% of the time, replace with random word else: masked_token = choice(vocab_list) masked_token_labels.append(tokens[index]) # Once we've saved the true label for that token, we can overwrite it with the masked version tokens[index] = masked_token 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, vocab_list): """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_chunk.append(segment) 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 = randint(1, len(current_chunk) - 1) tokens_a = [] for j in range(a_end): tokens_a.extend(current_chunk[j]) tokens_b = [] # Random next if len(current_chunk) == 1 or 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 = randint(0, len(random_document) - 1) for j in range(random_start, len(random_document)): tokens_b.extend(random_document[j]) if len(tokens_b) >= target_b_length: break # 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 else: is_random_next = False for j in range(a_end, len(current_chunk)): tokens_b.extend(current_chunk[j]) truncate_seq_pair(tokens_a, tokens_b, max_num_tokens) assert len(tokens_a) >= 1 assert len(tokens_b) >= 1 tokens = ["[CLS]"] + tokens_a + ["[SEP]"] + tokens_b + ["[SEP]"] # 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, masked_lm_positions, masked_lm_labels = create_masked_lm_predictions( tokens, masked_lm_prob, max_predictions_per_seq, vocab_list) 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} instances.append(instance) current_chunk = [] current_length = 0 i += 1 return instances def main(): parser = ArgumentParser() parser.add_argument('--train_corpus', type=Path, required=True) 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-base-cased", "bert-base-multilingual", "bert-base-chinese"]) parser.add_argument("--do_lower_case", action="store_true") 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") args = parser.parse_args() # TODO Add a low-memory / multiprocessing path for very large datasets # In this path documents would be stored in a shelf after being tokenized, and multiple processes would convert # those docs into training examples that would be written out on the fly. This would avoid the need to keep # the whole training set in memory and would speed up dataset creation at the cost of code complexity. # In addition, the finetuning script would need to be modified # to store the training epochs as memmapped arrays. tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) vocab_list = list(tokenizer.vocab.keys()) with args.train_corpus.open() as f: docs = [] doc = [] for line in tqdm(f, desc="Loading Dataset"): line = line.strip() if line == "": docs.append(doc) doc = [] else: tokens = tokenizer.tokenize(line) doc.append(tokens) args.output_dir.mkdir(exist_ok=True) docs = DocumentDatabase(docs) # When choosing a random sentence, we should sample docs proportionally to the number of sentences they contain # Google BERT doesn't do this, and as a result oversamples shorter docs for epoch in trange(args.epochs_to_generate, desc="Epoch"): epoch_filename = args.output_dir / f"epoch_{epoch}.json" num_instances = 0 with epoch_filename.open('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, vocab_list=vocab_list) 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 / f"epoch_{epoch}_metrics.json" with metrics_file.open('w') as metrics_file: metrics = { "num_training_examples": num_instances, "max_seq_len": args.max_seq_len } metrics_file.write(json.dumps(metrics)) if __name__ == '__main__': main()