Merge branch 'xlnet'

examples/utils_squad_evaluate.py

+""" Official evaluation script for SQuAD version 2.0.
+    Modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
+
+In addition to basic functionality, we also compute additional statistics and
+plot precision-recall curves if an additional na_prob.json file is provided.
+This file is expected to map question ID's to the model's predicted probability
+that a question is unanswerable.
+"""
+import argparse
+import collections
+import json
+import numpy as np
+import os
+import re
+import string
+import sys
+
+class EVAL_OPTS():
+  def __init__(self, data_file, pred_file, out_file="",
+               na_prob_file="na_prob.json", na_prob_thresh=1.0,
+               out_image_dir=None, verbose=False):
+    self.data_file = data_file
+    self.pred_file = pred_file
+    self.out_file = out_file
+    self.na_prob_file = na_prob_file
+    self.na_prob_thresh = na_prob_thresh
+    self.out_image_dir = out_image_dir
+    self.verbose = verbose
+
+OPTS = None
+
+def parse_args():
+  parser = argparse.ArgumentParser('Official evaluation script for SQuAD version 2.0.')
+  parser.add_argument('data_file', metavar='data.json', help='Input data JSON file.')
+  parser.add_argument('pred_file', metavar='pred.json', help='Model predictions.')
+  parser.add_argument('--out-file', '-o', metavar='eval.json',
+                      help='Write accuracy metrics to file (default is stdout).')
+  parser.add_argument('--na-prob-file', '-n', metavar='na_prob.json',
+                      help='Model estimates of probability of no answer.')
+  parser.add_argument('--na-prob-thresh', '-t', type=float, default=1.0,
+                      help='Predict "" if no-answer probability exceeds this (default = 1.0).')
+  parser.add_argument('--out-image-dir', '-p', metavar='out_images', default=None,
+                      help='Save precision-recall curves to directory.')
+  parser.add_argument('--verbose', '-v', action='store_true')
+  if len(sys.argv) == 1:
+    parser.print_help()
+    sys.exit(1)
+  return parser.parse_args()
+
+def make_qid_to_has_ans(dataset):
+  qid_to_has_ans = {}
+  for article in dataset:
+    for p in article['paragraphs']:
+      for qa in p['qas']:
+        qid_to_has_ans[qa['id']] = bool(qa['answers'])
+  return qid_to_has_ans
+
+def normalize_answer(s):
+  """Lower text and remove punctuation, articles and extra whitespace."""
+  def remove_articles(text):
+    regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
+    return re.sub(regex, ' ', text)
+  def white_space_fix(text):
+    return ' '.join(text.split())
+  def remove_punc(text):
+    exclude = set(string.punctuation)
+    return ''.join(ch for ch in text if ch not in exclude)
+  def lower(text):
+    return text.lower()
+  return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+def get_tokens(s):
+  if not s: return []
+  return normalize_answer(s).split()
+
+def compute_exact(a_gold, a_pred):
+  return int(normalize_answer(a_gold) == normalize_answer(a_pred))
+
+def compute_f1(a_gold, a_pred):
+  gold_toks = get_tokens(a_gold)
+  pred_toks = get_tokens(a_pred)
+  common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
+  num_same = sum(common.values())
+  if len(gold_toks) == 0 or len(pred_toks) == 0:
+    # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
+    return int(gold_toks == pred_toks)
+  if num_same == 0:
+    return 0
+  precision = 1.0 * num_same / len(pred_toks)
+  recall = 1.0 * num_same / len(gold_toks)
+  f1 = (2 * precision * recall) / (precision + recall)
+  return f1
+
+def get_raw_scores(dataset, preds):
+  exact_scores = {}
+  f1_scores = {}
+  for article in dataset:
+    for p in article['paragraphs']:
+      for qa in p['qas']:
+        qid = qa['id']
+        gold_answers = [a['text'] for a in qa['answers']
+                        if normalize_answer(a['text'])]
+        if not gold_answers:
+          # For unanswerable questions, only correct answer is empty string
+          gold_answers = ['']
+        if qid not in preds:
+          print('Missing prediction for %s' % qid)
+          continue
+        a_pred = preds[qid]
+        # Take max over all gold answers
+        exact_scores[qid] = max(compute_exact(a, a_pred) for a in gold_answers)
+        f1_scores[qid] = max(compute_f1(a, a_pred) for a in gold_answers)
+  return exact_scores, f1_scores
+
+def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
+  new_scores = {}
+  for qid, s in scores.items():
+    pred_na = na_probs[qid] > na_prob_thresh
+    if pred_na:
+      new_scores[qid] = float(not qid_to_has_ans[qid])
+    else:
+      new_scores[qid] = s
+  return new_scores
+
+def make_eval_dict(exact_scores, f1_scores, qid_list=None):
+  if not qid_list:
+    total = len(exact_scores)
+    return collections.OrderedDict([
+        ('exact', 100.0 * sum(exact_scores.values()) / total),
+        ('f1', 100.0 * sum(f1_scores.values()) / total),
+        ('total', total),
+    ])
+  else:
+    total = len(qid_list)
+    return collections.OrderedDict([
+        ('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
+        ('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
+        ('total', total),
+    ])
+
+def merge_eval(main_eval, new_eval, prefix):
+  for k in new_eval:
+    main_eval['%s_%s' % (prefix, k)] = new_eval[k]
+
+def plot_pr_curve(precisions, recalls, out_image, title):
+  plt.step(recalls, precisions, color='b', alpha=0.2, where='post')
+  plt.fill_between(recalls, precisions, step='post', alpha=0.2, color='b')
+  plt.xlabel('Recall')
+  plt.ylabel('Precision')
+  plt.xlim([0.0, 1.05])
+  plt.ylim([0.0, 1.05])
+  plt.title(title)
+  plt.savefig(out_image)
+  plt.clf()
+
+def make_precision_recall_eval(scores, na_probs, num_true_pos, qid_to_has_ans,
+                               out_image=None, title=None):
+  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+  true_pos = 0.0
+  cur_p = 1.0
+  cur_r = 0.0
+  precisions = [1.0]
+  recalls = [0.0]
+  avg_prec = 0.0
+  for i, qid in enumerate(qid_list):
+    if qid_to_has_ans[qid]:
+      true_pos += scores[qid]
+    cur_p = true_pos / float(i+1)
+    cur_r = true_pos / float(num_true_pos)
+    if i == len(qid_list) - 1 or na_probs[qid] != na_probs[qid_list[i+1]]:
+      # i.e., if we can put a threshold after this point
+      avg_prec += cur_p * (cur_r - recalls[-1])
+      precisions.append(cur_p)
+      recalls.append(cur_r)
+  if out_image:
+    plot_pr_curve(precisions, recalls, out_image, title)
+  return {'ap': 100.0 * avg_prec}
+
+def run_precision_recall_analysis(main_eval, exact_raw, f1_raw, na_probs, 
+                                  qid_to_has_ans, out_image_dir):
+  if out_image_dir and not os.path.exists(out_image_dir):
+    os.makedirs(out_image_dir)
+  num_true_pos = sum(1 for v in qid_to_has_ans.values() if v)
+  if num_true_pos == 0:
+    return
+  pr_exact = make_precision_recall_eval(
+      exact_raw, na_probs, num_true_pos, qid_to_has_ans,
+      out_image=os.path.join(out_image_dir, 'pr_exact.png'),
+      title='Precision-Recall curve for Exact Match score')
+  pr_f1 = make_precision_recall_eval(
+      f1_raw, na_probs, num_true_pos, qid_to_has_ans,
+      out_image=os.path.join(out_image_dir, 'pr_f1.png'),
+      title='Precision-Recall curve for F1 score')
+  oracle_scores = {k: float(v) for k, v in qid_to_has_ans.items()}
+  pr_oracle = make_precision_recall_eval(
+      oracle_scores, na_probs, num_true_pos, qid_to_has_ans,
+      out_image=os.path.join(out_image_dir, 'pr_oracle.png'),
+      title='Oracle Precision-Recall curve (binary task of HasAns vs. NoAns)')
+  merge_eval(main_eval, pr_exact, 'pr_exact')
+  merge_eval(main_eval, pr_f1, 'pr_f1')
+  merge_eval(main_eval, pr_oracle, 'pr_oracle')
+
+def histogram_na_prob(na_probs, qid_list, image_dir, name):
+  if not qid_list:
+    return
+  x = [na_probs[k] for k in qid_list]
+  weights = np.ones_like(x) / float(len(x))
+  plt.hist(x, weights=weights, bins=20, range=(0.0, 1.0))
+  plt.xlabel('Model probability of no-answer')
+  plt.ylabel('Proportion of dataset')
+  plt.title('Histogram of no-answer probability: %s' % name)
+  plt.savefig(os.path.join(image_dir, 'na_prob_hist_%s.png' % name))
+  plt.clf()
+
+def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
+  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+  cur_score = num_no_ans
+  best_score = cur_score
+  best_thresh = 0.0
+  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+  for i, qid in enumerate(qid_list):
+    if qid not in scores: continue
+    if qid_to_has_ans[qid]:
+      diff = scores[qid]
+    else:
+      if preds[qid]:
+        diff = -1
+      else:
+        diff = 0
+    cur_score += diff
+    if cur_score > best_score:
+      best_score = cur_score
+      best_thresh = na_probs[qid]
+  return 100.0 * best_score / len(scores), best_thresh
+
+def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
+  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+  cur_score = num_no_ans
+  best_score = cur_score
+  best_thresh = 0.0
+  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+  for i, qid in enumerate(qid_list):
+    if qid not in scores: continue
+    if qid_to_has_ans[qid]:
+      diff = scores[qid]
+    else:
+      if preds[qid]:
+        diff = -1
+      else:
+        diff = 0
+    cur_score += diff
+    if cur_score > best_score:
+      best_score = cur_score
+      best_thresh = na_probs[qid]
+
+  has_ans_score, has_ans_cnt = 0, 0
+  for qid in qid_list:
+    if not qid_to_has_ans[qid]: continue
+    has_ans_cnt += 1
+
+    if qid not in scores: continue
+    has_ans_score += scores[qid]
+
+  return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
+
+def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+  best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
+  best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
+  main_eval['best_exact'] = best_exact
+  main_eval['best_exact_thresh'] = exact_thresh
+  main_eval['best_f1'] = best_f1
+  main_eval['best_f1_thresh'] = f1_thresh
+
+def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+  best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(preds, exact_raw, na_probs, qid_to_has_ans)
+  best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(preds, f1_raw, na_probs, qid_to_has_ans)
+  main_eval['best_exact'] = best_exact
+  main_eval['best_exact_thresh'] = exact_thresh
+  main_eval['best_f1'] = best_f1
+  main_eval['best_f1_thresh'] = f1_thresh
+  main_eval['has_ans_exact'] = has_ans_exact
+  main_eval['has_ans_f1'] = has_ans_f1
+
+def main(OPTS):
+  with open(OPTS.data_file) as f:
+    dataset_json = json.load(f)
+    dataset = dataset_json['data']
+  with open(OPTS.pred_file) as f:
+    preds = json.load(f)
+  if OPTS.na_prob_file:
+    with open(OPTS.na_prob_file) as f:
+      na_probs = json.load(f)
+  else:
+    na_probs = {k: 0.0 for k in preds}
+  qid_to_has_ans = make_qid_to_has_ans(dataset)  # maps qid to True/False
+  has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
+  no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
+  exact_raw, f1_raw = get_raw_scores(dataset, preds)
+  exact_thresh = apply_no_ans_threshold(exact_raw, na_probs, qid_to_has_ans,
+                                        OPTS.na_prob_thresh)
+  f1_thresh = apply_no_ans_threshold(f1_raw, na_probs, qid_to_has_ans,
+                                     OPTS.na_prob_thresh)
+  out_eval = make_eval_dict(exact_thresh, f1_thresh)
+  if has_ans_qids:
+    has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
+    merge_eval(out_eval, has_ans_eval, 'HasAns')
+  if no_ans_qids:
+    no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
+    merge_eval(out_eval, no_ans_eval, 'NoAns')
+  if OPTS.na_prob_file:
+    find_all_best_thresh(out_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans)
+  if OPTS.na_prob_file and OPTS.out_image_dir:
+    run_precision_recall_analysis(out_eval, exact_raw, f1_raw, na_probs, 
+                                  qid_to_has_ans, OPTS.out_image_dir)
+    histogram_na_prob(na_probs, has_ans_qids, OPTS.out_image_dir, 'hasAns')
+    histogram_na_prob(na_probs, no_ans_qids, OPTS.out_image_dir, 'noAns')
+  if OPTS.out_file:
+    with open(OPTS.out_file, 'w') as f:
+      json.dump(out_eval, f)
+  else:
+    print(json.dumps(out_eval, indent=2))
+  return out_eval
+
+if __name__ == '__main__':
+  OPTS = parse_args()
+  if OPTS.out_image_dir:
+    import matplotlib
+    matplotlib.use('Agg')
+    import matplotlib.pyplot as plt 
+  main(OPTS)

hubconfs/bert_hubconf.py

-from pytorch_pretrained_bert.tokenization import BertTokenizer
-from pytorch_pretrained_bert.modeling import (
+from pytorch_transformers.tokenization_bert import BertTokenizer
+from pytorch_transformers.modeling_bert import (
         BertModel,
         BertForNextSentencePrediction,
         BertForMaskedLM,
@@ -86,7 +86,7 @@ def bertTokenizer(*args, **kwargs):
     Example:
         >>> import torch
         >>> sentence = 'Hello, World!'
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         >>> toks = tokenizer.tokenize(sentence)
         ['Hello', '##,', 'World', '##!']
         >>> ids = tokenizer.convert_tokens_to_ids(toks)
@@ -106,7 +106,7 @@ def bertModel(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -115,7 +115,7 @@ def bertModel(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertModel', 'bert-base-cased')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertModel', 'bert-base-cased')
         >>> model.eval()
         # Predict hidden states features for each layer
         >>> with torch.no_grad():
@@ -135,7 +135,7 @@ def bertForNextSentencePrediction(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -144,7 +144,7 @@ def bertForNextSentencePrediction(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertForNextSentencePrediction
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForNextSentencePrediction', 'bert-base-cased')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForNextSentencePrediction', 'bert-base-cased')
         >>> model.eval()
         # Predict the next sentence classification logits
         >>> with torch.no_grad():
@@ -165,7 +165,7 @@ def bertForPreTraining(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -173,7 +173,7 @@ def bertForPreTraining(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertForPreTraining
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForPreTraining', 'bert-base-cased')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForPreTraining', 'bert-base-cased')
         >>> masked_lm_logits_scores, seq_relationship_logits = model(tokens_tensor, segments_tensors)
     """
     model = BertForPreTraining.from_pretrained(*args, **kwargs)
@@ -189,7 +189,7 @@ def bertForMaskedLM(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -200,7 +200,7 @@ def bertForMaskedLM(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertForMaskedLM
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForMaskedLM', 'bert-base-cased')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForMaskedLM', 'bert-base-cased')
         >>> model.eval()
         # Predict all tokens
         >>> with torch.no_grad():
@@ -231,7 +231,7 @@ def bertForSequenceClassification(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -240,7 +240,7 @@ def bertForSequenceClassification(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertForSequenceClassification
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForSequenceClassification', 'bert-base-cased', num_labels=2)
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForSequenceClassification', 'bert-base-cased', num_labels=2)
         >>> model.eval()
         # Predict the sequence classification logits
         >>> with torch.no_grad():
@@ -266,7 +266,7 @@ def bertForMultipleChoice(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -275,7 +275,7 @@ def bertForMultipleChoice(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens, indexed_tokens]).unsqueeze(0)
         >>> segments_tensors = torch.tensor([segments_ids, segments_ids]).unsqueeze(0)
         # Load bertForMultipleChoice
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForMultipleChoice', 'bert-base-cased', num_choices=2)
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForMultipleChoice', 'bert-base-cased', num_choices=2)
         >>> model.eval()
         # Predict the multiple choice logits
         >>> with torch.no_grad():
@@ -299,7 +299,7 @@ def bertForQuestionAnswering(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -308,7 +308,7 @@ def bertForQuestionAnswering(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertForQuestionAnswering
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForQuestionAnswering', 'bert-base-cased')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForQuestionAnswering', 'bert-base-cased')
         >>> model.eval()
         # Predict the start and end positions logits
         >>> with torch.no_grad():
@@ -338,7 +338,7 @@ def bertForTokenClassification(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
         #  Prepare tokenized input
         >>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -347,7 +347,7 @@ def bertForTokenClassification(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
         >>> segments_tensors = torch.tensor([segments_ids])
         # Load bertForTokenClassification
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForTokenClassification', 'bert-base-cased', num_labels=2)
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'bertForTokenClassification', 'bert-base-cased', num_labels=2)
         >>> model.eval()
         # Predict the token classification logits
         >>> with torch.no_grad():

hubconfs/gpt2_hubconf.py

-from pytorch_pretrained_bert.tokenization_gpt2 import GPT2Tokenizer
-from pytorch_pretrained_bert.modeling_gpt2 import (
+from pytorch_transformers.tokenization_gpt2 import GPT2Tokenizer
+from pytorch_transformers.modeling_gpt2 import (
     GPT2Model,
     GPT2LMHeadModel,
     GPT2DoubleHeadsModel
@@ -53,7 +53,7 @@ def gpt2Tokenizer(*args, **kwargs):
 
     Example:
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2Tokenizer', 'gpt2')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'gpt2Tokenizer', 'gpt2')
 
         >>> text = "Who was Jim Henson ?"
         >>> indexed_tokens = tokenizer.encode(tokenized_text)
@@ -72,7 +72,7 @@ def gpt2Model(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2Tokenizer', 'gpt2')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'gpt2Tokenizer', 'gpt2')
 
         #  Prepare tokenized input
         >>> text_1 = "Who was Jim Henson ?"
@@ -83,7 +83,7 @@ def gpt2Model(*args, **kwargs):
         >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
 
         # Load gpt2Model
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2Model', 'gpt2')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'gpt2Model', 'gpt2')
         >>> model.eval()
 
         # Predict hidden states features for each layer
@@ -105,7 +105,7 @@ def gpt2LMHeadModel(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2Tokenizer', 'gpt2')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'gpt2Tokenizer', 'gpt2')
 
         #  Prepare tokenized input
         >>> text_1 = "Who was Jim Henson ?"
@@ -116,7 +116,7 @@ def gpt2LMHeadModel(*args, **kwargs):
         >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
 
         # Load gpt2LMHeadModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2LMHeadModel', 'gpt2')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'gpt2LMHeadModel', 'gpt2')
         >>> model.eval()
 
         # Predict hidden states features for each layer
@@ -144,7 +144,7 @@ def gpt2DoubleHeadsModel(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2Tokenizer', 'gpt2')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'gpt2Tokenizer', 'gpt2')
 
         #  Prepare tokenized input
         >>> text1 = "Who was Jim Henson ? Jim Henson was a puppeteer"
@@ -157,7 +157,7 @@ def gpt2DoubleHeadsModel(*args, **kwargs):
         >>> mc_token_ids = torch.LongTensor([[len(tokenized_text1)-1, len(tokenized_text2)-1]])
 
         # Load gpt2DoubleHeadsModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'gpt2DoubleHeadsModel', 'gpt2')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'gpt2DoubleHeadsModel', 'gpt2')
         >>> model.eval()
 
         # Predict hidden states features for each layer

hubconfs/gpt_hubconf.py

-from pytorch_pretrained_bert.tokenization_openai import OpenAIGPTTokenizer
-from pytorch_pretrained_bert.modeling_openai import (
+from pytorch_transformers.tokenization_openai import OpenAIGPTTokenizer
+from pytorch_transformers.modeling_openai import (
 	OpenAIGPTModel,
 	OpenAIGPTLMHeadModel,
 	OpenAIGPTDoubleHeadsModel
@@ -77,7 +77,7 @@ def openAIGPTTokenizer(*args, **kwargs):
 
     Example:
 		>>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTTokenizer', 'openai-gpt')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTTokenizer', 'openai-gpt')
 		
 		>>> text = "Who was Jim Henson ? Jim Henson was a puppeteer"
         >>> tokenized_text = tokenizer.tokenize(text)
@@ -98,7 +98,7 @@ def openAIGPTModel(*args, **kwargs):
     Example:
         # Load the tokenizer
 		>>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTTokenizer', 'openai-gpt')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTTokenizer', 'openai-gpt')
 
         #  Prepare tokenized input
         >>> text = "Who was Jim Henson ? Jim Henson was a puppeteer"
@@ -107,7 +107,7 @@ def openAIGPTModel(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
 
         # Load openAIGPTModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTModel', 'openai-gpt')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTModel', 'openai-gpt')
         >>> model.eval()
 
         # Predict hidden states features for each layer
@@ -127,7 +127,7 @@ def openAIGPTLMHeadModel(*args, **kwargs):
 	Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTTokenizer', 'openai-gpt')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTTokenizer', 'openai-gpt')
 
         #  Prepare tokenized input
         >>> text = "Who was Jim Henson ? Jim Henson was a puppeteer"
@@ -136,7 +136,7 @@ def openAIGPTLMHeadModel(*args, **kwargs):
         >>> tokens_tensor = torch.tensor([indexed_tokens])
 
         # Load openAIGPTLMHeadModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTLMHeadModel', 'openai-gpt')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTLMHeadModel', 'openai-gpt')
         >>> model.eval()
 
         # Predict hidden states features for each layer
@@ -162,7 +162,7 @@ def openAIGPTDoubleHeadsModel(*args, **kwargs):
 	Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTTokenizer', 'openai-gpt')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTTokenizer', 'openai-gpt')
 
         #  Prepare tokenized input
         >>> text1 = "Who was Jim Henson ? Jim Henson was a puppeteer"
@@ -175,7 +175,7 @@ def openAIGPTDoubleHeadsModel(*args, **kwargs):
         >>> mc_token_ids = torch.LongTensor([[len(tokenized_text1)-1, len(tokenized_text2)-1]])
 
         # Load openAIGPTDoubleHeadsModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'openAIGPTDoubleHeadsModel', 'openai-gpt')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'openAIGPTDoubleHeadsModel', 'openai-gpt')
         >>> model.eval()
 
         # Predict hidden states features for each layer

hubconfs/transformer_xl_hubconf.py

-from pytorch_pretrained_bert.tokenization_transfo_xl import TransfoXLTokenizer
-from pytorch_pretrained_bert.modeling_transfo_xl import (
+from pytorch_transformers.tokenization_transfo_xl import TransfoXLTokenizer
+from pytorch_transformers.modeling_transfo_xl import (
     TransfoXLModel,
     TransfoXLLMHeadModel
 )
@@ -46,7 +46,7 @@ def transformerXLTokenizer(*args, **kwargs):
 
     Example:
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'transformerXLTokenizer', 'transfo-xl-wt103')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'transformerXLTokenizer', 'transfo-xl-wt103')
         
         >>> text = "Who was Jim Henson ?"
         >>> tokenized_text = tokenizer.tokenize(tokenized_text)
@@ -64,7 +64,7 @@ def transformerXLModel(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'transformerXLTokenizer', 'transfo-xl-wt103')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'transformerXLTokenizer', 'transfo-xl-wt103')
 
         #  Prepare tokenized input
         >>> text_1 = "Who was Jim Henson ?"
@@ -77,7 +77,7 @@ def transformerXLModel(*args, **kwargs):
         >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
 
         # Load transformerXLModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'transformerXLModel', 'transfo-xl-wt103')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'transformerXLModel', 'transfo-xl-wt103')
         >>> model.eval()
 
         # Predict hidden states features for each layer
@@ -99,7 +99,7 @@ def transformerXLLMHeadModel(*args, **kwargs):
     Example:
         # Load the tokenizer
         >>> import torch
-        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'transformerXLTokenizer', 'transfo-xl-wt103')
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'transformerXLTokenizer', 'transfo-xl-wt103')
 
         #  Prepare tokenized input
         >>> text_1 = "Who was Jim Henson ?"
@@ -112,7 +112,7 @@ def transformerXLLMHeadModel(*args, **kwargs):
         >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
 
         # Load transformerXLLMHeadModel
-        >>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'transformerXLLMHeadModel', 'transfo-xl-wt103')
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'transformerXLLMHeadModel', 'transfo-xl-wt103')
         >>> model.eval()
 
         # Predict hidden states features for each layer

hubconfs/xlm_hubconf.py

+from pytorch_transformers.tokenization_xlm import XLMTokenizer
+from pytorch_transformers.modeling_xlm import (
+    XLMConfig,
+    XLMModel,
+    XLMWithLMHeadModel,
+    XLMForSequenceClassification,
+    XLMForQuestionAnswering
+)
+
+# A lot of models share the same param doc. Use a decorator
+# to save typing
+xlm_start_docstring = """
+    Model class adapted from the XLM Transformer model of
+        "Cross-lingual Language Model Pretraining" by Guillaume Lample, Alexis Conneau
+        Paper: https://arxiv.org/abs/1901.07291
+        Original code: https://github.com/facebookresearch/XLM
+
+    Example:
+        # Load the tokenizer
+        >>> import torch
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlmTokenizer', 'xlm-mlm-en-2048')
+
+        #  Prepare tokenized input
+        >>> text_1 = "Who was Jim Henson ?"
+        >>> text_2 = "Jim Henson was a puppeteer"
+        >>> indexed_tokens_1 = tokenizer.encode(text_1)
+        >>> indexed_tokens_2 = tokenizer.encode(text_2)
+        >>> tokens_tensor_1 = torch.tensor([indexed_tokens_1])
+        >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
+"""
+
+# A lot of models share the same param doc. Use a decorator
+# to save typing
+xlm_end_docstring = """
+    Params:
+        pretrained_model_name_or_path: either:
+            - a str with the name of a pre-trained model to load selected in the list of:
+                . `xlm-mlm-en-2048`
+            - a path or url to a pretrained model archive containing:
+                . `config.json` a configuration file for the model
+                . `pytorch_model.bin` a PyTorch dump created using the `convert_xlm_checkpoint_to_pytorch` conversion script
+        cache_dir: an optional path to a folder in which the pre-trained models will be cached.
+        state_dict: an optional state dictionary (collections.OrderedDict object) to use instead of pre-trained models
+        *inputs, **kwargs: additional input for the specific XLM class
+"""
+
+
+def _begin_with_docstring(docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = fn.__doc__ + docstr
+        return fn
+    return docstring_decorator
+
+def _end_with_docstring(docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = fn.__doc__ + docstr
+        return fn
+    return docstring_decorator
+
+
+def xlmTokenizer(*args, **kwargs):
+    """
+    Instantiate a XLM BPE tokenizer for XLM from a pre-trained vocab file.
+
+    Args:
+    pretrained_model_name_or_path: Path to pretrained model archive
+                                   or one of pre-trained vocab configs below.
+                                       * xlm-mlm-en-2048
+    Keyword args:
+    special_tokens: Special tokens in vocabulary that are not pretrained
+                    Default: None
+    max_len: An artificial maximum length to truncate tokenized sequences to;
+             Effective maximum length is always the minimum of this
+             value (if specified) and the underlying model's
+             sequence length.
+             Default: None
+
+    Example:
+        >>> import torch
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlmTokenizer', 'xlm-mlm-en-2048')
+
+        >>> text = "Who was Jim Henson ?"
+        >>> indexed_tokens = tokenizer.encode(tokenized_text)
+    """
+    tokenizer = XLMTokenizer.from_pretrained(*args, **kwargs)
+    return tokenizer
+
+
+@_begin_with_docstring(xlm_start_docstring)
+@_end_with_docstring(xlm_end_docstring)
+def xlmModel(*args, **kwargs):
+    """
+        # Load xlmModel
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'xlmModel', 'xlm-mlm-en-2048')
+        >>> model.eval()
+
+        # Predict hidden states features for each layer
+        >>> with torch.no_grad():
+                hidden_states_1, mems = model(tokens_tensor_1)
+                hidden_states_2, mems = model(tokens_tensor_2, past=mems)
+    """
+    model = XLMModel.from_pretrained(*args, **kwargs)
+    return model
+
+
+@_begin_with_docstring(xlm_start_docstring)
+@_end_with_docstring(xlm_end_docstring)
+def xlmLMHeadModel(*args, **kwargs):
+    """
+        #  Prepare tokenized input
+        >>> text_1 = "Who was Jim Henson ?"
+        >>> text_2 = "Jim Henson was a puppeteer"
+        >>> indexed_tokens_1 = tokenizer.encode(text_1)
+        >>> indexed_tokens_2 = tokenizer.encode(text_2)
+        >>> tokens_tensor_1 = torch.tensor([indexed_tokens_1])
+        >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
+
+        # Load xlnetLMHeadModel
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'xlnetLMHeadModel', 'xlm-mlm-en-2048')
+        >>> model.eval()
+
+        # Predict hidden states features for each layer
+        >>> with torch.no_grad():
+                predictions_1, mems = model(tokens_tensor_1)
+                predictions_2, mems = model(tokens_tensor_2, mems=mems)
+
+        # Get the predicted last token
+        >>> predicted_index = torch.argmax(predictions_2[0, -1, :]).item()
+        >>> predicted_token = tokenizer.decode([predicted_index])
+        >>> assert predicted_token == ' who'
+    """
+    model = XLMWithLMHeadModel.from_pretrained(*args, **kwargs)
+    return model
+
+
+# @_end_with_docstring(xlnet_docstring)
+# def xlnetForSequenceClassification(*args, **kwargs):
+#     """
+#     xlnetModel is the basic XLNet Transformer model from
+#         "XLNet: Generalized Autoregressive Pretraining for Language Understanding"
+#         by Zhilin Yang, Zihang Dai1, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le
+
+#     Example:
+#         # Load the tokenizer
+#         >>> import torch
+#         >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlnetTokenizer', 'xlm-mlm-en-2048')
+
+#         #  Prepare tokenized input
+#         >>> text1 = "Who was Jim Henson ? Jim Henson was a puppeteer"
+#         >>> text2 = "Who was Jim Henson ? Jim Henson was a mysterious young man"
+#         >>> tokenized_text1 = tokenizer.tokenize(text1)
+#         >>> tokenized_text2 = tokenizer.tokenize(text2)
+#         >>> indexed_tokens1 = tokenizer.convert_tokens_to_ids(tokenized_text1)
+#         >>> indexed_tokens2 = tokenizer.convert_tokens_to_ids(tokenized_text2)
+#         >>> tokens_tensor = torch.tensor([[indexed_tokens1, indexed_tokens2]])
+#         >>> mc_token_ids = torch.LongTensor([[len(tokenized_text1)-1, len(tokenized_text2)-1]])
+
+#         # Load xlnetForSequenceClassification
+#         >>> model = torch.hub.load('huggingface/pytorch-transformers', 'xlnetForSequenceClassification', 'xlm-mlm-en-2048')
+#         >>> model.eval()
+
+#         # Predict sequence classes logits
+#         >>> with torch.no_grad():
+#                 lm_logits, mems = model(tokens_tensor)
+#     """
+#     model = XLNetForSequenceClassification.from_pretrained(*args, **kwargs)
+#     return model

hubconfs/xlnet_hubconf.1.py

+from pytorch_transformers.tokenization_xlnet import XLNetTokenizer
+from pytorch_transformers.modeling_xlnet import (
+    XLNetConfig,
+    XLNetModel,
+    XLNetLMHeadModel,
+    # XLNetForSequenceClassification
+)
+
+# A lot of models share the same param doc. Use a decorator
+# to save typing
+xlnet_docstring = """
+    Params:
+        pretrained_model_name_or_path: either:
+            - a str with the name of a pre-trained model to load selected in the list of:
+                . `xlnet-large-cased`
+            - a path or url to a pretrained model archive containing:
+                . `config.json` a configuration file for the model
+                . `pytorch_model.bin` a PyTorch dump of a XLNetForPreTraining instance
+            - a path or url to a pretrained model archive containing:
+                . `xlnet_config.json` a configuration file for the model
+                . `model.chkpt` a TensorFlow checkpoint
+        from_tf: should we load the weights from a locally saved TensorFlow checkpoint
+        cache_dir: an optional path to a folder in which the pre-trained models will be cached.
+        state_dict: an optional state dictionary (collections.OrderedDict object) to use instead of pre-trained models
+        *inputs, **kwargs: additional input for the specific XLNet class
+"""
+
+
+def _append_from_pretrained_docstring(docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = fn.__doc__ + docstr
+        return fn
+    return docstring_decorator
+
+
+def xlnetTokenizer(*args, **kwargs):
+    """
+    Instantiate a XLNet sentencepiece tokenizer for XLNet from a pre-trained vocab file.
+    Peculiarities:
+        - require Google sentencepiece (https://github.com/google/sentencepiece)
+
+    Args:
+    pretrained_model_name_or_path: Path to pretrained model archive
+                                   or one of pre-trained vocab configs below.
+                                       * xlnet-large-cased
+    Keyword args:
+    special_tokens: Special tokens in vocabulary that are not pretrained
+                    Default: None
+    max_len: An artificial maximum length to truncate tokenized sequences to;
+             Effective maximum length is always the minimum of this
+             value (if specified) and the underlying model's
+             sequence length.
+             Default: None
+
+    Example:
+        >>> import torch
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlnetTokenizer', 'xlnet-large-cased')
+
+        >>> text = "Who was Jim Henson ?"
+        >>> indexed_tokens = tokenizer.encode(tokenized_text)
+    """
+    tokenizer = XLNetTokenizer.from_pretrained(*args, **kwargs)
+    return tokenizer
+
+
+@_append_from_pretrained_docstring(xlnet_docstring)
+def xlnetModel(*args, **kwargs):
+    """
+    xlnetModel is the basic XLNet Transformer model from
+        "XLNet: Generalized Autoregressive Pretraining for Language Understanding"
+        by Zhilin Yang, Zihang Dai1, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le
+
+    Example:
+        # Load the tokenizer
+        >>> import torch
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlnetTokenizer', 'xlnet-large-cased')
+
+        #  Prepare tokenized input
+        >>> text_1 = "Who was Jim Henson ?"
+        >>> text_2 = "Jim Henson was a puppeteer"
+        >>> indexed_tokens_1 = tokenizer.encode(text_1)
+        >>> indexed_tokens_2 = tokenizer.encode(text_2)
+        >>> tokens_tensor_1 = torch.tensor([indexed_tokens_1])
+        >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
+
+        # Load xlnetModel
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'xlnetModel', 'xlnet-large-cased')
+        >>> model.eval()
+
+        # Predict hidden states features for each layer
+        >>> with torch.no_grad():
+                hidden_states_1, mems = model(tokens_tensor_1)
+                hidden_states_2, mems = model(tokens_tensor_2, past=mems)
+    """
+    model = XLNetModel.from_pretrained(*args, **kwargs)
+    return model
+
+
+@_append_from_pretrained_docstring(xlnet_docstring)
+def xlnetLMHeadModel(*args, **kwargs):
+    """
+    xlnetModel is the basic XLNet Transformer model from
+        "XLNet: Generalized Autoregressive Pretraining for Language Understanding"
+        by Zhilin Yang, Zihang Dai1, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le
+    with a tied (pre-trained) language modeling head on top.
+
+    Example:
+        # Load the tokenizer
+        >>> import torch
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlnetTokenizer', 'xlnet-large-cased')
+
+        #  Prepare tokenized input
+        >>> text_1 = "Who was Jim Henson ?"
+        >>> text_2 = "Jim Henson was a puppeteer"
+        >>> indexed_tokens_1 = tokenizer.encode(text_1)
+        >>> indexed_tokens_2 = tokenizer.encode(text_2)
+        >>> tokens_tensor_1 = torch.tensor([indexed_tokens_1])
+        >>> tokens_tensor_2 = torch.tensor([indexed_tokens_2])
+
+        # Load xlnetLMHeadModel
+        >>> model = torch.hub.load('huggingface/pytorch-transformers', 'xlnetLMHeadModel', 'xlnet-large-cased')
+        >>> model.eval()
+
+        # Predict hidden states features for each layer
+        >>> with torch.no_grad():
+                predictions_1, mems = model(tokens_tensor_1)
+                predictions_2, mems = model(tokens_tensor_2, mems=mems)
+
+        # Get the predicted last token
+        >>> predicted_index = torch.argmax(predictions_2[0, -1, :]).item()
+        >>> predicted_token = tokenizer.decode([predicted_index])
+        >>> assert predicted_token == ' who'
+    """
+    model = XLNetLMHeadModel.from_pretrained(*args, **kwargs)
+    return model
+
+
+# @_append_from_pretrained_docstring(xlnet_docstring)
+# def xlnetForSequenceClassification(*args, **kwargs):
+#     """
+#     xlnetModel is the basic XLNet Transformer model from
+#         "XLNet: Generalized Autoregressive Pretraining for Language Understanding"
+#         by Zhilin Yang, Zihang Dai1, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le
+
+#     Example:
+#         # Load the tokenizer
+#         >>> import torch
+#         >>> tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'xlnetTokenizer', 'xlnet-large-cased')
+
+#         #  Prepare tokenized input
+#         >>> text1 = "Who was Jim Henson ? Jim Henson was a puppeteer"
+#         >>> text2 = "Who was Jim Henson ? Jim Henson was a mysterious young man"
+#         >>> tokenized_text1 = tokenizer.tokenize(text1)
+#         >>> tokenized_text2 = tokenizer.tokenize(text2)
+#         >>> indexed_tokens1 = tokenizer.convert_tokens_to_ids(tokenized_text1)
+#         >>> indexed_tokens2 = tokenizer.convert_tokens_to_ids(tokenized_text2)
+#         >>> tokens_tensor = torch.tensor([[indexed_tokens1, indexed_tokens2]])
+#         >>> mc_token_ids = torch.LongTensor([[len(tokenized_text1)-1, len(tokenized_text2)-1]])
+
+#         # Load xlnetForSequenceClassification
+#         >>> model = torch.hub.load('huggingface/pytorch-transformers', 'xlnetForSequenceClassification', 'xlnet-large-cased')
+#         >>> model.eval()
+
+#         # Predict sequence classes logits
+#         >>> with torch.no_grad():
+#                 lm_logits, mems = model(tokens_tensor)
+#     """
+#     model = XLNetForSequenceClassification.from_pretrained(*args, **kwargs)
+#     return model

notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb

@@ -78,7 +78,7 @@
     "import importlib.util\n",
     "import sys\n",
     "import tensorflow as tf\n",
-    "import pytorch_pretrained_bert as ppb\n",
+    "import pytorch_transformers as ppb\n",
     "\n",
     "def del_all_flags(FLAGS):\n",
     "    flags_dict = FLAGS._flags()    \n",
@@ -3997,9 +3997,9 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "11/16/2018 11:03:05 - INFO - pytorch_pretrained_bert.modeling -   loading archive file https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz from cache at /Users/thomaswolf/.pytorch_pretrained_bert/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba\n",
-      "11/16/2018 11:03:05 - INFO - pytorch_pretrained_bert.modeling -   extracting archive file /Users/thomaswolf/.pytorch_pretrained_bert/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba to temp dir /var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmpaqgsm566\n",
-      "11/16/2018 11:03:08 - INFO - pytorch_pretrained_bert.modeling -   Model config {\n",
+      "11/16/2018 11:03:05 - INFO - pytorch_transformers.modeling_bert -   loading archive file https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz from cache at /Users/thomaswolf/.pytorch_transformers/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba\n",
+      "11/16/2018 11:03:05 - INFO - pytorch_transformers.modeling_bert -   extracting archive file /Users/thomaswolf/.pytorch_transformers/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba to temp dir /var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmpaqgsm566\n",
+      "11/16/2018 11:03:08 - INFO - pytorch_transformers.modeling_bert -   Model config {\n",
       "  \"attention_probs_dropout_prob\": 0.1,\n",
       "  \"hidden_act\": \"gelu\",\n",
       "  \"hidden_dropout_prob\": 0.1,\n",

notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb

@@ -86,7 +86,7 @@
     "spec.loader.exec_module(module)\n",
     "sys.modules['modeling_tensorflow'] = module\n",
     "\n",
-    "spec = importlib.util.spec_from_file_location('*', original_tf_inplem_dir + '/run_squad.py')\n",
+    "spec = importlib.util.spec_from_file_location('*', original_tf_inplem_dir + '/run_bert_squad.py')\n",
     "module = importlib.util.module_from_spec(spec)\n",
     "spec.loader.exec_module(module)\n",
     "sys.modules['run_squad_tensorflow'] = module\n",

notebooks/Comparing-TF-and-PT-models.ipynb

@@ -342,7 +342,7 @@
    "outputs": [],
    "source": [
     "import extract_features\n",
-    "import pytorch_pretrained_bert as ppb\n",
+    "import pytorch_transformers as ppb\n",
     "from extract_features import *"
    ]
   },
@@ -375,8 +375,8 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "11/15/2018 16:21:18 - INFO - pytorch_pretrained_bert.modeling -   loading archive file ../../google_models/uncased_L-12_H-768_A-12/\n",
-      "11/15/2018 16:21:18 - INFO - pytorch_pretrained_bert.modeling -   Model config {\n",
+      "11/15/2018 16:21:18 - INFO - pytorch_transformers.modeling_bert -   loading archive file ../../google_models/uncased_L-12_H-768_A-12/\n",
+      "11/15/2018 16:21:18 - INFO - pytorch_transformers.modeling_bert -   Model config {\n",
       "  \"attention_probs_dropout_prob\": 0.1,\n",
       "  \"hidden_act\": \"gelu\",\n",
       "  \"hidden_dropout_prob\": 0.1,\n",

pytorch_pretrained_bert/__init__.py

-__version__ = "0.6.2"
-from .tokenization import BertTokenizer, BasicTokenizer, WordpieceTokenizer
-from .tokenization_openai import OpenAIGPTTokenizer
-from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
-from .tokenization_gpt2 import GPT2Tokenizer
-
-from .modeling import (BertConfig, BertModel, BertForPreTraining,
-                       BertForMaskedLM, BertForNextSentencePrediction,
-                       BertForSequenceClassification, BertForMultipleChoice,
-                       BertForTokenClassification, BertForQuestionAnswering,
-                       load_tf_weights_in_bert)
-from .modeling_openai import (OpenAIGPTConfig, OpenAIGPTModel,
-                              OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
-                              load_tf_weights_in_openai_gpt)
-from .modeling_transfo_xl import (TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel,
-                                  load_tf_weights_in_transfo_xl)
-from .modeling_gpt2 import (GPT2Config, GPT2Model,
-                            GPT2LMHeadModel, GPT2DoubleHeadsModel, GPT2MultipleChoiceHead,
-                            load_tf_weights_in_gpt2)
-
-from .optimization import BertAdam
-from .optimization_openai import OpenAIAdam
-
-from .file_utils import PYTORCH_PRETRAINED_BERT_CACHE, cached_path, WEIGHTS_NAME, CONFIG_NAME

pytorch_pretrained_bert/__main__.py

-# coding: utf8
-def main():
-    import sys
-    if (len(sys.argv) != 4 and len(sys.argv) != 5) or sys.argv[1] not in [
-        "convert_tf_checkpoint_to_pytorch",
-        "convert_openai_checkpoint",
-        "convert_transfo_xl_checkpoint",
-        "convert_gpt2_checkpoint",
-    ]:
-        print(
-        "Should be used as one of: \n"
-        ">> `pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`, \n"
-        ">> `pytorch_pretrained_bert convert_openai_checkpoint OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG]`, \n"
-        ">> `pytorch_pretrained_bert convert_transfo_xl_checkpoint TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG]` or \n"
-        ">> `pytorch_pretrained_bert convert_gpt2_checkpoint TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG]`")
-    else:
-        if sys.argv[1] == "convert_tf_checkpoint_to_pytorch":
-            try:
-                from .convert_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
-            except ImportError:
-                print("pytorch_pretrained_bert can only be used from the commandline to convert TensorFlow models in PyTorch, "
-                    "In that case, it requires TensorFlow to be installed. Please see "
-                    "https://www.tensorflow.org/install/ for installation instructions.")
-                raise
-
-            if len(sys.argv) != 5:
-                # pylint: disable=line-too-long
-                print("Should be used as `pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
-            else:
-                PYTORCH_DUMP_OUTPUT = sys.argv.pop()
-                TF_CONFIG = sys.argv.pop()
-                TF_CHECKPOINT = sys.argv.pop()
-                convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
-        elif sys.argv[1] == "convert_openai_checkpoint":
-            from .convert_openai_checkpoint_to_pytorch import convert_openai_checkpoint_to_pytorch
-            OPENAI_GPT_CHECKPOINT_FOLDER_PATH = sys.argv[2]
-            PYTORCH_DUMP_OUTPUT = sys.argv[3]
-            if len(sys.argv) == 5:
-                OPENAI_GPT_CONFIG = sys.argv[4]
-            else:
-                OPENAI_GPT_CONFIG = ""
-            convert_openai_checkpoint_to_pytorch(OPENAI_GPT_CHECKPOINT_FOLDER_PATH,
-                                                 OPENAI_GPT_CONFIG,
-                                                 PYTORCH_DUMP_OUTPUT)
-        elif sys.argv[1] == "convert_transfo_xl_checkpoint":
-            try:
-                from .convert_transfo_xl_checkpoint_to_pytorch import convert_transfo_xl_checkpoint_to_pytorch
-            except ImportError:
-                print("pytorch_pretrained_bert can only be used from the commandline to convert TensorFlow models in PyTorch, "
-                    "In that case, it requires TensorFlow to be installed. Please see "
-                    "https://www.tensorflow.org/install/ for installation instructions.")
-                raise
-
-            if 'ckpt' in sys.argv[2].lower():
-                TF_CHECKPOINT = sys.argv[2]
-                TF_DATASET_FILE = ""
-            else:
-                TF_DATASET_FILE = sys.argv[2]
-                TF_CHECKPOINT = ""
-            PYTORCH_DUMP_OUTPUT = sys.argv[3]
-            if len(sys.argv) == 5:
-                TF_CONFIG = sys.argv[4]
-            else:
-                TF_CONFIG = ""
-            convert_transfo_xl_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT, TF_DATASET_FILE)
-        else:
-            try:
-                from .convert_gpt2_checkpoint_to_pytorch import convert_gpt2_checkpoint_to_pytorch
-            except ImportError:
-                print("pytorch_pretrained_bert can only be used from the commandline to convert TensorFlow models in PyTorch, "
-                    "In that case, it requires TensorFlow to be installed. Please see "
-                    "https://www.tensorflow.org/install/ for installation instructions.")
-                raise
-
-            TF_CHECKPOINT = sys.argv[2]
-            PYTORCH_DUMP_OUTPUT = sys.argv[3]
-            if len(sys.argv) == 5:
-                TF_CONFIG = sys.argv[4]
-            else:
-                TF_CONFIG = ""
-            convert_gpt2_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
-if __name__ == '__main__':
-    main()

pytorch_pretrained_bert/optimization.py

-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""PyTorch optimization for BERT model."""
-
-import math
-import torch
-from torch.optim import Optimizer
-from torch.optim.optimizer import required
-from torch.nn.utils import clip_grad_norm_
-import logging
-import abc
-import sys
-
-logger = logging.getLogger(__name__)
-
-
-if sys.version_info >= (3, 4):
-    ABC = abc.ABC
-else:
-    ABC = abc.ABCMeta('ABC', (), {})
-
-
-class _LRSchedule(ABC):
-    """ Parent of all LRSchedules here. """
-    warn_t_total = False        # is set to True for schedules where progressing beyond t_total steps doesn't make sense
-    def __init__(self, warmup=0.002, t_total=-1, **kw):
-        """
-        :param warmup:  what fraction of t_total steps will be used for linear warmup
-        :param t_total: how many training steps (updates) are planned
-        :param kw:
-        """
-        super(_LRSchedule, self).__init__(**kw)
-        if t_total < 0:
-            logger.warning("t_total value of {} results in schedule not being applied".format(t_total))
-        if not 0.0 <= warmup < 1.0 and not warmup == -1:
-            raise ValueError("Invalid warmup: {} - should be in [0.0, 1.0[ or -1".format(warmup))
-        warmup = max(warmup, 0.)
-        self.warmup, self.t_total = float(warmup), float(t_total)
-        self.warned_for_t_total_at_progress = -1
-
-    def get_lr(self, step, nowarn=False):
-        """
-        :param step:    which of t_total steps we're on
-        :param nowarn:  set to True to suppress warning regarding training beyond specified 't_total' steps
-        :return:        learning rate multiplier for current update
-        """
-        if self.t_total < 0:
-            return 1.
-        progress = float(step) / self.t_total
-        ret = self.get_lr_(progress)
-        # warning for exceeding t_total (only active with warmup_linear
-        if not nowarn and self.warn_t_total and progress > 1. and progress > self.warned_for_t_total_at_progress:
-            logger.warning(
-                "Training beyond specified 't_total'. Learning rate multiplier set to {}. Please set 't_total' of {} correctly."
-                    .format(ret, self.__class__.__name__))
-            self.warned_for_t_total_at_progress = progress
-        # end warning
-        return ret
-
-    @abc.abstractmethod
-    def get_lr_(self, progress):
-        """
-        :param progress:    value between 0 and 1 (unless going beyond t_total steps) specifying training progress
-        :return:            learning rate multiplier for current update
-        """
-        return 1.
-
-
-class ConstantLR(_LRSchedule):
-    def get_lr_(self, progress):
-        return 1.
-
-
-class WarmupCosineSchedule(_LRSchedule):
-    """
-    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
-    Decreases learning rate from 1. to 0. over remaining `1 - warmup` steps following a cosine curve.
-    If `cycles` (default=0.5) is different from default, learning rate follows cosine function after warmup.
-    """
-    warn_t_total = True
-    def __init__(self, warmup=0.002, t_total=-1, cycles=.5, **kw):
-        """
-        :param warmup:      see LRSchedule
-        :param t_total:     see LRSchedule
-        :param cycles:      number of cycles. Default: 0.5, corresponding to cosine decay from 1. at progress==warmup and 0 at progress==1.
-        :param kw:
-        """
-        super(WarmupCosineSchedule, self).__init__(warmup=warmup, t_total=t_total, **kw)
-        self.cycles = cycles
-
-    def get_lr_(self, progress):
-        if progress < self.warmup:
-            return progress / self.warmup
-        else:
-            progress = (progress - self.warmup) / (1 - self.warmup)   # progress after warmup
-            return 0.5 * (1. + math.cos(math.pi * self.cycles * 2 * progress))
-
-
-class WarmupCosineWithHardRestartsSchedule(WarmupCosineSchedule):
-    """
-    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
-    If `cycles` (default=1.) is different from default, learning rate follows `cycles` times a cosine decaying
-    learning rate (with hard restarts).
-    """
-    def __init__(self, warmup=0.002, t_total=-1, cycles=1., **kw):
-        super(WarmupCosineWithHardRestartsSchedule, self).__init__(warmup=warmup, t_total=t_total, cycles=cycles, **kw)
-        assert(cycles >= 1.)
-
-    def get_lr_(self, progress):
-        if progress < self.warmup:
-            return progress / self.warmup
-        else:
-            progress = (progress - self.warmup) / (1 - self.warmup)     # progress after warmup
-            ret = 0.5 * (1. + math.cos(math.pi * ((self.cycles * progress) % 1)))
-            return ret
-
-
-class WarmupCosineWithWarmupRestartsSchedule(WarmupCosineWithHardRestartsSchedule):
-    """
-    All training progress is divided in `cycles` (default=1.) parts of equal length.
-    Every part follows a schedule with the first `warmup` fraction of the training steps linearly increasing from 0. to 1.,
-    followed by a learning rate decreasing from 1. to 0. following a cosine curve.
-    """
-    def __init__(self, warmup=0.002, t_total=-1, cycles=1., **kw):
-        assert(warmup * cycles < 1.)
-        warmup = warmup * cycles if warmup >= 0 else warmup
-        super(WarmupCosineWithWarmupRestartsSchedule, self).__init__(warmup=warmup, t_total=t_total, cycles=cycles, **kw)
-
-    def get_lr_(self, progress):
-        progress = progress * self.cycles % 1.
-        if progress < self.warmup:
-            return progress / self.warmup
-        else:
-            progress = (progress - self.warmup) / (1 - self.warmup)     # progress after warmup
-            ret = 0.5 * (1. + math.cos(math.pi * progress))
-            return ret
-
-
-class WarmupConstantSchedule(_LRSchedule):
-    """
-    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
-    Keeps learning rate equal to 1. after warmup.
-    """
-    def get_lr_(self, progress):
-        if progress < self.warmup:
-            return progress / self.warmup
-        return 1.
-
-
-class WarmupLinearSchedule(_LRSchedule):
-    """
-    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
-    Linearly decreases learning rate from 1. to 0. over remaining `1 - warmup` steps.
-    """
-    warn_t_total = True
-    def get_lr_(self, progress):
-        if progress < self.warmup:
-            return progress / self.warmup
-        return max((progress - 1.) / (self.warmup - 1.), 0.)
-
-
-SCHEDULES = {
-    None:       ConstantLR,
-    "none":     ConstantLR,
-    "warmup_cosine": WarmupCosineSchedule,
-    "warmup_constant": WarmupConstantSchedule,
-    "warmup_linear": WarmupLinearSchedule
-}
-
-
-class BertAdam(Optimizer):
-    """Implements BERT version of Adam algorithm with weight decay fix.
-    Params:
-        lr: learning rate
-        warmup: portion of t_total for the warmup, -1  means no warmup. Default: -1
-        t_total: total number of training steps for the learning
-            rate schedule, -1  means constant learning rate of 1. (no warmup regardless of warmup setting). Default: -1
-        schedule: schedule to use for the warmup (see above).
-            Can be `'warmup_linear'`, `'warmup_constant'`, `'warmup_cosine'`, `'none'`, `None` or a `_LRSchedule` object (see below).
-            If `None` or `'none'`, learning rate is always kept constant.
-            Default : `'warmup_linear'`
-        betas: Adams betas. Default: (0.9, 0.999)
-        e: Adams epsilon. Default: 1e-6
-        weight_decay: Weight decay. Default: 0.01
-        max_grad_norm: Maximum norm for the gradients (-1 means no clipping). Default: 1.0
-    """
-    def __init__(self, params, lr=required, warmup=-1, t_total=-1, schedule='warmup_linear',
-                 betas=(0.9, 0.999), e=1e-6, weight_decay=0.01, max_grad_norm=1.0, **kwargs):
-        if lr is not required and lr < 0.0:
-            raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
-        if not isinstance(schedule, _LRSchedule) and schedule not in SCHEDULES:
-            raise ValueError("Invalid schedule parameter: {}".format(schedule))
-        if not 0.0 <= betas[0] < 1.0:
-            raise ValueError("Invalid beta parameter at index 0: {} - should be in [0.0, 1.0[".format(betas[0]))
-        if not 0.0 <= betas[1] < 1.0:
-            raise ValueError("Invalid beta parameter at index 1: {} - should be in [0.0, 1.0[".format(betas[1]))
-        if not e >= 0.0:
-            raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(e))
-        # initialize schedule object
-        if not isinstance(schedule, _LRSchedule):
-            schedule_type = SCHEDULES[schedule]
-            schedule = schedule_type(warmup=warmup, t_total=t_total)
-        else:
-            if warmup != -1 or t_total != -1:
-                logger.warning("warmup and t_total on the optimizer are ineffective when _LRSchedule object is provided as schedule. "
-                               "Please specify custom warmup and t_total in _LRSchedule object.")
-        defaults = dict(lr=lr, schedule=schedule,
-                        betas=betas, e=e, weight_decay=weight_decay,
-                        max_grad_norm=max_grad_norm)
-        super(BertAdam, self).__init__(params, defaults)
-
-    def get_lr(self):
-        lr = []
-        for group in self.param_groups:
-            for p in group['params']:
-                state = self.state[p]
-                if len(state) == 0:
-                    return [0]
-                lr_scheduled = group['lr']
-                lr_scheduled *= group['schedule'].get_lr(state['step'])
-                lr.append(lr_scheduled)
-        return lr
-
-    def step(self, closure=None):
-        """Performs a single optimization step.
-
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
-        loss = None
-        if closure is not None:
-            loss = closure()
-
-        for group in self.param_groups:
-            for p in group['params']:
-                if p.grad is None:
-                    continue
-                grad = p.grad.data
-                if grad.is_sparse:
-                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
-
-                state = self.state[p]
-
-                # State initialization
-                if len(state) == 0:
-                    state['step'] = 0
-                    # Exponential moving average of gradient values
-                    state['next_m'] = torch.zeros_like(p.data)
-                    # Exponential moving average of squared gradient values
-                    state['next_v'] = torch.zeros_like(p.data)
-
-                next_m, next_v = state['next_m'], state['next_v']
-                beta1, beta2 = group['betas']
-
-                # Add grad clipping
-                if group['max_grad_norm'] > 0:
-                    clip_grad_norm_(p, group['max_grad_norm'])
-
-                # Decay the first and second moment running average coefficient
-                # In-place operations to update the averages at the same time
-                next_m.mul_(beta1).add_(1 - beta1, grad)
-                next_v.mul_(beta2).addcmul_(1 - beta2, grad, grad)
-                update = next_m / (next_v.sqrt() + group['e'])
-
-                # Just adding the square of the weights to the loss function is *not*
-                # the correct way of using L2 regularization/weight decay with Adam,
-                # since that will interact with the m and v parameters in strange ways.
-                #
-                # Instead we want to decay the weights in a manner that doesn't interact
-                # with the m/v parameters. This is equivalent to adding the square
-                # of the weights to the loss with plain (non-momentum) SGD.
-                if group['weight_decay'] > 0.0:
-                    update += group['weight_decay'] * p.data
-
-                lr_scheduled = group['lr']
-                lr_scheduled *= group['schedule'].get_lr(state['step'])
-
-                update_with_lr = lr_scheduled * update
-                p.data.add_(-update_with_lr)
-
-                state['step'] += 1
-
-                # step_size = lr_scheduled * math.sqrt(bias_correction2) / bias_correction1
-                # No bias correction
-                # bias_correction1 = 1 - beta1 ** state['step']
-                # bias_correction2 = 1 - beta2 ** state['step']
-
-        return loss

pytorch_pretrained_bert/optimization_openai.py

-# coding=utf-8
-# Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""PyTorch optimization for OpenAI GPT model."""
-
-import math
-import torch
-from torch.optim import Optimizer
-from torch.optim.optimizer import required
-from torch.nn.utils import clip_grad_norm_
-import logging
-from .optimization import SCHEDULES, _LRSchedule, WarmupCosineWithWarmupRestartsSchedule, \
-    WarmupCosineWithHardRestartsSchedule, WarmupCosineSchedule, WarmupLinearSchedule, WarmupConstantSchedule
-
-logger = logging.getLogger(__name__)
-
-
-class OpenAIAdam(Optimizer):
-    """Implements Open AI version of Adam algorithm with weight decay fix.
-    """
-    def __init__(self, params, lr=required, schedule='warmup_linear', warmup=-1, t_total=-1,
-                 betas=(0.9, 0.999), e=1e-8, weight_decay=0,
-                 vector_l2=False, max_grad_norm=-1, **kwargs):
-        if lr is not required and lr < 0.0:
-            raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
-        if not isinstance(schedule, _LRSchedule) and schedule not in SCHEDULES:
-            raise ValueError("Invalid schedule parameter: {}".format(schedule))
-        if not 0.0 <= betas[0] < 1.0:
-            raise ValueError("Invalid beta parameter at index 0: {} - should be in [0.0, 1.0[".format(betas[0]))
-        if not 0.0 <= betas[1] < 1.0:
-            raise ValueError("Invalid beta parameter at index 1: {} - should be in [0.0, 1.0[".format(betas[1]))
-        if not e >= 0.0:
-            raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(e))
-        # initialize schedule object
-        if not isinstance(schedule, _LRSchedule):
-            schedule_type = SCHEDULES[schedule]
-            schedule = schedule_type(warmup=warmup, t_total=t_total)
-        else:
-            if warmup != -1 or t_total != -1:
-                logger.warning("warmup and t_total on the optimizer are ineffective when _LRSchedule object is provided as schedule. "
-                               "Please specify custom warmup and t_total in _LRSchedule object.")
-        defaults = dict(lr=lr, schedule=schedule,
-                        betas=betas, e=e, weight_decay=weight_decay, vector_l2=vector_l2,
-                        max_grad_norm=max_grad_norm)
-        super(OpenAIAdam, self).__init__(params, defaults)
-
-    def get_lr(self):
-        lr = []
-        for group in self.param_groups:
-            for p in group['params']:
-                state = self.state[p]
-                if len(state) == 0:
-                    return [0]
-                lr_scheduled = group['lr']
-                lr_scheduled *= group['schedule'].get_lr(state['step'])
-                lr.append(lr_scheduled)
-        return lr
-
-    def step(self, closure=None):
-        """Performs a single optimization step.
-
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
-        loss = None
-        if closure is not None:
-            loss = closure()
-
-        for group in self.param_groups:
-            for p in group['params']:
-                if p.grad is None:
-                    continue
-                grad = p.grad.data
-                if grad.is_sparse:
-                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
-
-                state = self.state[p]
-
-                # State initialization
-                if len(state) == 0:
-                    state['step'] = 0
-                    # Exponential moving average of gradient values
-                    state['exp_avg'] = torch.zeros_like(p.data)
-                    # Exponential moving average of squared gradient values
-                    state['exp_avg_sq'] = torch.zeros_like(p.data)
-
-                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
-                beta1, beta2 = group['betas']
-
-                state['step'] += 1
-
-                # Add grad clipping
-                if group['max_grad_norm'] > 0:
-                    clip_grad_norm_(p, group['max_grad_norm'])
-
-                # Decay the first and second moment running average coefficient
-                exp_avg.mul_(beta1).add_(1 - beta1, grad)
-                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
-                denom = exp_avg_sq.sqrt().add_(group['e'])
-
-                bias_correction1 = 1 - beta1 ** state['step']
-                bias_correction2 = 1 - beta2 ** state['step']
-
-                lr_scheduled = group['lr']
-                lr_scheduled *= group['schedule'].get_lr(state['step'])
-
-                step_size = lr_scheduled * math.sqrt(bias_correction2) / bias_correction1
-
-                p.data.addcdiv_(-step_size, exp_avg, denom)
-
-                # Add weight decay at the end (fixed version)
-                if (len(p.size()) > 1 or group['vector_l2']) and group['weight_decay'] > 0:
-                    p.data.add_(-lr_scheduled * group['weight_decay'], p.data)
-
-        return loss

pytorch_transformers/__init__.py

+__version__ = "0.7.0"
+from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
+from .tokenization_openai import OpenAIGPTTokenizer
+from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
+from .tokenization_gpt2 import GPT2Tokenizer
+from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
+from .tokenization_xlm import XLMTokenizer
+from .tokenization_utils import (PreTrainedTokenizer, clean_up_tokenization)
+
+from .modeling_bert import (BertConfig, BertModel, BertForPreTraining,
+                       BertForMaskedLM, BertForNextSentencePrediction,
+                       BertForSequenceClassification, BertForMultipleChoice,
+                       BertForTokenClassification, BertForQuestionAnswering,
+                       load_tf_weights_in_bert, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+                       BERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
+from .modeling_openai import (OpenAIGPTConfig, OpenAIGPTModel,
+                              OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
+                              load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                              OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_transfo_xl import (TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel,
+                                  load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_gpt2 import (GPT2Config, GPT2Model,
+                            GPT2LMHeadModel, GPT2DoubleHeadsModel,
+                            load_tf_weights_in_gpt2, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                            GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_xlnet import (XLNetConfig,
+                             XLNetPreTrainedModel, XLNetModel, XLNetLMHeadModel,
+                             XLNetForSequenceClassification, XLNetForQuestionAnswering,
+                             load_tf_weights_in_xlnet, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                             XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_xlm import (XLMConfig, XLMModel,
+                           XLMWithLMHeadModel, XLMForSequenceClassification,
+                           XLMForQuestionAnswering, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                           XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_utils import (WEIGHTS_NAME, CONFIG_NAME, TF_WEIGHTS_NAME,
+                          PretrainedConfig, PreTrainedModel, prune_layer, Conv1D)
+
+from .optimization import (AdamW, ConstantLRSchedule, WarmupConstantSchedule, WarmupCosineSchedule,
+                           WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)
+
+from .file_utils import (PYTORCH_PRETRAINED_BERT_CACHE, cached_path)

pytorch_transformers/__main__.py

+# coding: utf8
+def main():
+    import sys
+    if (len(sys.argv) < 4 or len(sys.argv) > 6) or sys.argv[1] not in ["bert", "gpt", "transfo_xl", "gpt2", "xlnet", "xlm"]:
+        print(
+        "Should be used as one of: \n"
+        ">> pytorch_transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
+        ">> pytorch_transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
+        ">> pytorch_transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
+        ">> pytorch_transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
+        ">> pytorch_transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME] or \n"
+        ">> pytorch_transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT")
+    else:
+        if sys.argv[1] == "bert":
+            try:
+                from .convert_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
+            if len(sys.argv) != 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
+            else:
+                PYTORCH_DUMP_OUTPUT = sys.argv.pop()
+                TF_CONFIG = sys.argv.pop()
+                TF_CHECKPOINT = sys.argv.pop()
+                convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "gpt":
+            from .convert_openai_checkpoint_to_pytorch import convert_openai_checkpoint_to_pytorch
+            if len(sys.argv) < 4 or len(sys.argv) > 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG]`")
+            else:
+                OPENAI_GPT_CHECKPOINT_FOLDER_PATH = sys.argv[2]
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+                if len(sys.argv) == 5:
+                    OPENAI_GPT_CONFIG = sys.argv[4]
+                else:
+                    OPENAI_GPT_CONFIG = ""
+                convert_openai_checkpoint_to_pytorch(OPENAI_GPT_CHECKPOINT_FOLDER_PATH,
+                                                    OPENAI_GPT_CONFIG,
+                                                    PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "transfo_xl":
+            try:
+                from .convert_transfo_xl_checkpoint_to_pytorch import convert_transfo_xl_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+            if len(sys.argv) < 4 or len(sys.argv) > 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers transfo_xl TF_CHECKPOINT/TF_DATASET_FILE PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
+            else:
+                if 'ckpt' in sys.argv[2].lower():
+                    TF_CHECKPOINT = sys.argv[2]
+                    TF_DATASET_FILE = ""
+                else:
+                    TF_DATASET_FILE = sys.argv[2]
+                    TF_CHECKPOINT = ""
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+                if len(sys.argv) == 5:
+                    TF_CONFIG = sys.argv[4]
+                else:
+                    TF_CONFIG = ""
+                convert_transfo_xl_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT, TF_DATASET_FILE)
+        elif sys.argv[1] == "gpt2":
+            try:
+                from .convert_gpt2_checkpoint_to_pytorch import convert_gpt2_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
+            if len(sys.argv) < 4 or len(sys.argv) > 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
+            else:
+                TF_CHECKPOINT = sys.argv[2]
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+                if len(sys.argv) == 5:
+                    TF_CONFIG = sys.argv[4]
+                else:
+                    TF_CONFIG = ""
+                convert_gpt2_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "xlnet":
+            try:
+                from .convert_xlnet_checkpoint_to_pytorch import convert_xlnet_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
+            if len(sys.argv) < 5 or len(sys.argv) > 6:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME]`")
+            else:
+                TF_CHECKPOINT = sys.argv[2]
+                TF_CONFIG = sys.argv[3]
+                PYTORCH_DUMP_OUTPUT = sys.argv[4]
+                if len(sys.argv) == 6:
+                    FINETUNING_TASK = sys.argv[5]
+                else:
+                    FINETUNING_TASK = None
+
+                convert_xlnet_checkpoint_to_pytorch(TF_CHECKPOINT,
+                                                    TF_CONFIG,
+                                                    PYTORCH_DUMP_OUTPUT,
+                                                    FINETUNING_TASK)
+        elif sys.argv[1] == "xlm":
+            from .convert_xlm_checkpoint_to_pytorch import convert_xlm_checkpoint_to_pytorch
+
+            if len(sys.argv) != 4:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT`")
+            else:
+                XLM_CHECKPOINT_PATH = sys.argv[2]
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+
+                convert_xlm_checkpoint_to_pytorch(XLM_CHECKPOINT_PATH, PYTORCH_DUMP_OUTPUT)
+
+if __name__ == '__main__':
+    main()

pytorch_pretrained_bert/convert_gpt2_checkpoint_to_pytorch.py → pytorch_transformers/convert_gpt2_checkpoint_to_pytorch.py

@@ -21,11 +21,14 @@ from io import open
 
 import torch
 
-from pytorch_pretrained_bert.modeling_gpt2 import (CONFIG_NAME, WEIGHTS_NAME,
+from pytorch_transformers.modeling_gpt2 import (CONFIG_NAME, WEIGHTS_NAME,
                                                      GPT2Config,
                                                      GPT2Model,
                                                      load_tf_weights_in_gpt2)
 
+import logging
+logging.basicConfig(level=logging.INFO)
+
 
 def convert_gpt2_checkpoint_to_pytorch(gpt2_checkpoint_path, gpt2_config_file, pytorch_dump_folder_path):
     # Construct model
@@ -36,7 +39,7 @@ def convert_gpt2_checkpoint_to_pytorch(gpt2_checkpoint_path, gpt2_config_file, p
     model = GPT2Model(config)
 
     # Load weights from numpy
-    load_tf_weights_in_gpt2(model, gpt2_checkpoint_path)
+    load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path)
 
     # Save pytorch-model
     pytorch_weights_dump_path = pytorch_dump_folder_path + '/' + WEIGHTS_NAME

pytorch_pretrained_bert/convert_openai_checkpoint_to_pytorch.py → pytorch_transformers/convert_openai_checkpoint_to_pytorch.py

@@ -21,11 +21,14 @@ from io import open
 
 import torch
 
-from pytorch_pretrained_bert.modeling_openai import (CONFIG_NAME, WEIGHTS_NAME,
+from pytorch_transformers.modeling_openai import (CONFIG_NAME, WEIGHTS_NAME,
                                                      OpenAIGPTConfig,
                                                      OpenAIGPTModel,
                                                      load_tf_weights_in_openai_gpt)
 
+import logging
+logging.basicConfig(level=logging.INFO)
+
 
 def convert_openai_checkpoint_to_pytorch(openai_checkpoint_folder_path, openai_config_file, pytorch_dump_folder_path):
     # Construct model
@@ -36,7 +39,7 @@ def convert_openai_checkpoint_to_pytorch(openai_checkpoint_folder_path, openai_c
     model = OpenAIGPTModel(config)
 
     # Load weights from numpy
-    load_tf_weights_in_openai_gpt(model, openai_checkpoint_folder_path)
+    load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path)
 
     # Save pytorch-model
     pytorch_weights_dump_path = pytorch_dump_folder_path + '/' + WEIGHTS_NAME

pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py → pytorch_transformers/convert_tf_checkpoint_to_pytorch.py

@@ -18,14 +18,13 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import os
-import re
 import argparse
-import tensorflow as tf
 import torch
-import numpy as np
 
-from pytorch_pretrained_bert.modeling import BertConfig, BertForPreTraining, load_tf_weights_in_bert
+from pytorch_transformers.modeling_bert import BertConfig, BertForPreTraining, load_tf_weights_in_bert
+
+import logging
+logging.basicConfig(level=logging.INFO)
 
 def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytorch_dump_path):
     # Initialise PyTorch model
@@ -34,7 +33,7 @@ def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytor
     model = BertForPreTraining(config)
 
     # Load weights from tf checkpoint
-    load_tf_weights_in_bert(model, tf_checkpoint_path)
+    load_tf_weights_in_bert(model, config, tf_checkpoint_path)
 
     # Save pytorch-model
     print("Save PyTorch model to {}".format(pytorch_dump_path))