Added a --reduce_memory option to shelve docs to disc instead of keeping them in memory.

examples/lm_finetuning/pregenerate_training_data.py

@@ -1,46 +1,81 @@
 from argparse import ArgumentParser
 from pathlib import Path
 from tqdm import tqdm, trange
+from tempfile import TemporaryDirectory
+import shelve
 
 from random import random, randint, shuffle, choice, sample
 from pytorch_pretrained_bert.tokenization import BertTokenizer
-
+import numpy as np
 import json
 
 
 class DocumentDatabase:
-    def __init__(self, document_list):
+    def __init__(self, reduce_memory=False, working_dir=None):
-        self.document_list = document_list
+        if reduce_memory:
-        self.doc_starts = {}
+            if working_dir is None:
-        self.weighted_doc_samples = []
+                self.temp_dir = TemporaryDirectory()
-        i = 0
+                self.working_dir = Path(self.temp_dir.name)
-        for doc_idx, doc in enumerate(document_list):
+            else:
-            self.doc_starts[doc_idx] = i
+                self.temp_dir = None
-            self.weighted_doc_samples.extend([doc_idx] * len(doc))
+                self.working_dir = Path(working_dir)
-            i += len(doc)
+                self.working_dir.mkdir(parents=True, exist_ok=True)
+            self.document_shelf_filepath = self.working_dir / 'shelf.db'
+            self.document_shelf = shelve.open(str(self.document_shelf_filepath),
+                                              flag='n', protocol=-1)
+            self.documents = None
+        else:
+            self.documents = []
+            self.document_shelf = None
+            self.document_shelf_filepath = None
+        self.doc_lengths = []
+        self.doc_cumsum = None
+        self.cumsum_max = None
+        self.reduce_memory = reduce_memory
+
+    def add_document(self, document):
+        if self.reduce_memory:
+            current_idx = len(self.doc_lengths)
+            self.document_shelf[str(current_idx)] = document
+        else:
+            self.documents.append(document)
+        self.doc_lengths.append(len(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:
-            num_sentences = len(self.document_list[current_idx])
+            # With sentence weighting, we sample docs proportionally to their sentence length
-            # This very painful line randomly selects a document, weighted by the number of sentences they contain,
+            if self.doc_cumsum is None or len(self.doc_cumsum) != len(self.doc_lengths):
-            # while guaranteeing that it won't return the original document
+                self._precalculate_doc_weights()
-            sampled_val = (
+            rand_start = self.doc_cumsum[current_idx]
-                    (self.doc_starts[current_idx] + num_sentences
+            rand_end = rand_start + self.cumsum_max - self.doc_lengths[current_idx]
-                     + randint(0, len(self.weighted_doc_samples) - num_sentences - 1))
+            sentence_index = randint(rand_start, rand_end) % self.cumsum_max
-                    % len(self.weighted_doc_samples))
+            sampled_doc_index = np.searchsorted(self.doc_cumsum, sentence_index, side='right')
-            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)
+            sampled_doc_index = current_idx + randint(1, len(self.doc_lengths)-1)
         assert sampled_doc_index != current_idx
-        return self.document_list[sampled_doc_index]
+        if self.reduce_memory:
+            return self.document_shelf[str(sampled_doc_index)]
+        else:
+            return self.documents[sampled_doc_index]
 
     def __len__(self):
-        return len(self.document_list)
+        return len(self.doc_lengths)
 
     def __getitem__(self, item):
-        return self.document_list[item]
+        if self.reduce_memory:
+            return self.document_shelf[str(item)]
+        else:
+            return self.documents[item]
+
+    def cleanup(self):
+        if self.document_shelf is not None:
+            self.document_shelf.close()
 
 
 def truncate_seq_pair(tokens_a, tokens_b, max_num_tokens):
@@ -200,6 +235,11 @@ def main():
                                  "bert-base-multilingual", "bert-base-chinese"])
     parser.add_argument("--do_lower_case", action="store_true")
 
+    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("--working_dir", type=Path, default=None,
+                        help="Temporary directory to use for --reduce_memory. If not set, uses TemporaryDirectory()")
+
     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)
@@ -212,31 +252,21 @@ def main():
 
     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())
+    docs = DocumentDatabase(reduce_memory=args.reduce_memory, working_dir=args.working_dir)
     with args.train_corpus.open() as f:
-        docs = []
         doc = []
-        for line in tqdm(f, desc="Loading Dataset"):
+        for line in tqdm(f, desc="Loading Dataset", unit=" lines"):
             line = line.strip()
             if line == "":
-                docs.append(doc)
+                docs.add_document(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
@@ -257,6 +287,7 @@ def main():
                 "max_seq_len": args.max_seq_len
             }
             metrics_file.write(json.dumps(metrics))
+    docs.cleanup()
 
 
 if __name__ == '__main__':