added tests + fixed losses

pytorch_pretrained_bert/modeling.py

@@ -549,7 +549,7 @@ class BertPreTrainedModel(nn.Module):
                 model.__class__.__name__, unexpected_keys))
         if len(error_msgs) > 0:
             raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
-                               self.__class__.__name__, "\n\t".join(error_msgs)))
+                               model.__class__.__name__, "\n\t".join(error_msgs)))
         if tempdir:
             # Clean up temp dir
             shutil.rmtree(tempdir)

pytorch_pretrained_bert/tokenization_openai.py

@@ -67,19 +67,17 @@ class OpenAIGPTTokenizer(object):
     mostly a wrapper for a public python bpe tokenizer
     """
     @classmethod
-    def from_pretrained(cls, pretrained_model_name, cache_dir=None, *inputs, **kwargs):
+    def from_pretrained(cls, pretrained_model_name_or_path, cache_dir=None, *inputs, **kwargs):
         """
         Instantiate a PreTrainedBertModel from a pre-trained model file.
         Download and cache the pre-trained model file if needed.
         """
-        if pretrained_model_name in PRETRAINED_VOCAB_ARCHIVE_MAP:
-            vocab_file = PRETRAINED_VOCAB_ARCHIVE_MAP[pretrained_model_name]
-            merges_file = PRETRAINED_MERGES_ARCHIVE_MAP[pretrained_model_name]
+        if pretrained_model_name_or_path in PRETRAINED_VOCAB_ARCHIVE_MAP:
+            vocab_file = PRETRAINED_VOCAB_ARCHIVE_MAP[pretrained_model_name_or_path]
+            merges_file = PRETRAINED_MERGES_ARCHIVE_MAP[pretrained_model_name_or_path]
         else:
-            vocab_file = pretrained_model_name
-        if os.path.isdir(vocab_file):
-            vocab_file = os.path.join(vocab_file, VOCAB_NAME)
-            merges_file = os.path.join(vocab_file, MERGES_NAME)
+            vocab_file = os.path.join(pretrained_model_name_or_path, VOCAB_NAME)
+            merges_file = os.path.join(pretrained_model_name_or_path, MERGES_NAME)
         # redirect to the cache, if necessary
         try:
             resolved_vocab_file = cached_path(vocab_file, cache_dir=cache_dir)
@@ -87,11 +85,12 @@ class OpenAIGPTTokenizer(object):
         except FileNotFoundError:
             logger.error(
                 "Model name '{}' was not found in model name list ({}). "
-                "We assumed '{}' was a path or url but couldn't find any file "
-                "associated to this path or url.".format(
-                    pretrained_model_name,
+                "We assumed '{}' was a path or url but couldn't find files {} and {} "
+                "at this path or url.".format(
+                    pretrained_model_name_or_path,
                     ', '.join(PRETRAINED_VOCAB_ARCHIVE_MAP.keys()),
-                    vocab_file))
+                    pretrained_model_name_or_path,
+                    vocab_file, merges_file))
             return None
         if resolved_vocab_file == vocab_file and resolved_merges_file == merges_file:
             logger.info("loading vocabulary file {}".format(vocab_file))
@@ -101,29 +100,38 @@ class OpenAIGPTTokenizer(object):
                 vocab_file, resolved_vocab_file))
             logger.info("loading merges file {} from cache at {}".format(
                 merges_file, resolved_merges_file))
-        if pretrained_model_name in PRETRAINED_VOCAB_POSITIONAL_EMBEDDINGS_SIZE_MAP:
+        if pretrained_model_name_or_path in PRETRAINED_VOCAB_POSITIONAL_EMBEDDINGS_SIZE_MAP:
             # if we're using a pretrained model, ensure the tokenizer wont index sequences longer
             # than the number of positional embeddings
-            max_len = PRETRAINED_VOCAB_POSITIONAL_EMBEDDINGS_SIZE_MAP[pretrained_model_name]
+            max_len = PRETRAINED_VOCAB_POSITIONAL_EMBEDDINGS_SIZE_MAP[pretrained_model_name_or_path]
             kwargs['max_len'] = min(kwargs.get('max_len', int(1e12)), max_len)
         # Instantiate tokenizer.
         tokenizer = cls(resolved_vocab_file, resolved_merges_file, *inputs, **kwargs)
         return tokenizer
 
-    def __init__(self, vocab_file, merges_file):
+    def __init__(self, vocab_file, merges_file, special_tokens=None, max_len=None):
         try:
             import ftfy
             import spacy
         except ImportError:
             raise ImportError("Please install ftfy and spacy to use OpenAI GPT tokenizer.")
 
+        self.max_len = max_len if max_len is not None else int(1e12)
         self.nlp = spacy.load('en', disable=['parser', 'tagger', 'ner', 'textcat'])
+        self.fix_text = ftfy.fix_text
         self.encoder = json.load(open(vocab_file))
         self.decoder = {v:k for k,v in self.encoder.items()}
         merges = open(merges_file, encoding='utf-8').read().split('\n')[1:-1]
         merges = [tuple(merge.split()) for merge in merges]
         self.bpe_ranks = dict(zip(merges, range(len(merges))))
         self.cache = {}
+        if not special_tokens:
+            self.special_tokens = {}
+        else:
+            self.special_tokens = dict((tok, len(self.encoder) + i) for i, tok in enumerate(special_tokens))
+
+    def set_special_tokens(self, special_tokens):
+        self.special_tokens = dict((tok, len(self.encoder) + i) for i, tok in enumerate(special_tokens))
 
     def bpe(self, token):
         word = tuple(token[:-1]) + ( token[-1] + '</w>',)
@@ -168,20 +176,38 @@ class OpenAIGPTTokenizer(object):
         self.cache[token] = word
         return word
 
-    def tokenize(self, texts, verbose=True):
-        texts_tokens = []
-        if verbose:
-            for text in tqdm(texts, ncols=80, leave=False):
-                text = self.nlp(text_standardize(ftfy.fix_text(text)))
-                text_tokens = []
-                for token in text:
-                    text_tokens.extend([self.encoder.get(t, 0) for t in self.bpe(token.text.lower()).split(' ')])
-                texts_tokens.append(text_tokens)
-        else:
-            for text in texts:
-                text = self.nlp(text_standardize(ftfy.fix_text(text)))
-                text_tokens = []
-                for token in text:
-                    text_tokens.extend([self.encoder.get(t, 0) for t in self.bpe(token.text.lower()).split(' ')])
-                texts_tokens.append(text_tokens)
-        return texts_tokens
+    def tokenize(self, text):
+        split_tokens = []
+        text = self.nlp(text_standardize(self.fix_text(text)))
+        for token in text:
+            split_tokens.extend([t for t in self.bpe(token.text.lower()).split(' ')])
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        """Converts a sequence of tokens into ids using the vocab."""
+        ids = []
+        for token in tokens:
+            if token in self.special_tokens:
+                ids.append(self.special_tokens[token])
+            else:
+                ids.append(self.encoder.get(token, 0))
+        if len(ids) > self.max_len:
+            raise ValueError(
+                "Token indices sequence length is longer than the specified maximum "
+                " sequence length for this BERT model ({} > {}). Running this"
+                " sequence through BERT will result in indexing errors".format(len(ids), self.max_len)
+            )
+        return ids
+
+    def convert_ids_to_tokens(self, ids):
+        """Converts a sequence of ids in BPE tokens using the vocab."""
+        tokens = []
+        for i in ids:
+            tokens.append(self.decoder[i])
+        return tokens
+
+    def decode(self, ids):
+        """Converts a sequence of ids in a string."""
+        tokens = self.convert_ids_to_tokens(ids)
+        out_string = ''.join(tokens).replace('</w>', ' ')
+        return out_string

tests/modeling_openai_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import json
+import random
+
+import torch
+
+from pytorch_pretrained_bert import (OpenAIGPTConfig, OpenAIGPTModel, OpenAIGPTDoubleHeadsModel)
+
+
+class OpenAIGPTModelTest(unittest.TestCase):
+    class OpenAIGPTModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_position_ids=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     n_special=1,
+                     n_ctx=33,
+                     n_embd=32,
+                     n_layer=5,
+                     n_head=4,
+                     n_choices=3,
+                     afn="gelu",
+                     resid_pdrop=0.1,
+                     attn_pdrop=0.1,
+                     embd_pdrop=0.1,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     scope=None):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_position_ids = use_position_ids
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.n_special = n_special
+            self.n_ctx = n_ctx
+            self.n_embd = n_embd
+            self.n_layer = n_layer
+            self.n_head = n_head
+            self.afn = afn
+            self.n_choices = n_choices
+            self.resid_pdrop = resid_pdrop
+            self.attn_pdrop = attn_pdrop
+            self.embd_pdrop = embd_pdrop
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            input_ids = OpenAIGPTModelTest.ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.vocab_size)
+
+            position_ids = None
+            if self.use_position_ids:
+                position_ids = OpenAIGPTModelTest.ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.n_ctx)
+                position_ids = position_ids + self.n_special + self.vocab_size
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                total_voc = self.n_ctx + self.n_special + self.vocab_size
+                token_type_ids = OpenAIGPTModelTest.ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_voc)
+
+            multiple_choice_labels = None
+            lm_labels = None
+            classification_token_mask = None
+            if self.use_labels:
+                multiple_choice_labels = OpenAIGPTModelTest.ids_tensor([self.batch_size], self.type_sequence_label_size)
+                lm_labels = OpenAIGPTModelTest.ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.num_labels)
+                classification_token_mask = OpenAIGPTModelTest.ids_tensor([self.batch_size, self.n_choices, self.seq_length], 2).float()
+
+            config = OpenAIGPTConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                n_ctx=self.n_ctx,
+                n_special=self.n_special,
+                n_embd=self.n_embd,
+                n_layer=self.n_layer,
+                n_head=self.n_head,
+                afn=self.afn,
+                resid_pdrop=self.resid_pdrop,
+                attn_pdrop=self.attn_pdrop,
+                embd_pdrop=self.embd_pdrop,
+                initializer_range=self.initializer_range)
+
+            return (config, input_ids, token_type_ids, position_ids,
+                    multiple_choice_labels, lm_labels, classification_token_mask)
+
+        def create_openai_model(self, config, input_ids, token_type_ids, position_ids,
+                                multiple_choice_labels, lm_labels, classification_token_mask):
+            model = OpenAIGPTModel(config)
+            hidden_states = model(input_ids, position_ids, token_type_ids)
+            outputs = {
+                "hidden_states": hidden_states,
+            }
+            return outputs
+
+        def check_openai_model_output(self, result):
+            self.parent.assertListEqual(
+                list(result["hidden_states"].size()),
+                [self.batch_size, self.n_choices, self.seq_length, self.n_embd])
+
+
+        def create_openai_double_heads(self, config, input_ids, token_type_ids, position_ids,
+                                       multiple_choice_labels, lm_labels, classification_token_mask):
+            model = OpenAIGPTDoubleHeadsModel(config)
+            loss = model(input_ids, classification_token_mask, position_ids,
+                         token_type_ids, lm_labels, multiple_choice_labels)
+            lm_logits, multiple_choice_logits = model(input_ids, classification_token_mask, position_ids, token_type_ids)
+            outputs = {
+                "loss": loss,
+                "lm_logits": lm_logits,
+                "multiple_choice_logits": multiple_choice_logits,
+            }
+            return outputs
+
+        def check_openai_double_heads_output(self, result):
+            total_voc = self.n_ctx + self.n_special + self.vocab_size
+            self.parent.assertListEqual(
+                list(result["lm_logits"].size()),
+                [self.batch_size, self.n_choices, self.seq_length, total_voc])
+            self.parent.assertListEqual(
+                list(result["multiple_choice_logits"].size()),
+                [self.batch_size, self.n_choices])
+
+        def check_openai_double_heads_loss_output(self, result):
+            self.parent.assertListEqual(
+                [list(l.size()) for l in result["loss"]],
+                [[], []])
+
+    def test_default(self):
+        self.run_tester(OpenAIGPTModelTest.OpenAIGPTModelTester(self))
+
+    def test_config_to_json_string(self):
+        config = OpenAIGPTConfig(vocab_size_or_config_json_file=99, n_embd=37)
+        obj = json.loads(config.to_json_string())
+        self.assertEqual(obj["vocab_size"], 99)
+        self.assertEqual(obj["n_embd"], 37)
+
+    def run_tester(self, tester):
+        config_and_inputs = tester.prepare_config_and_inputs()
+        output_result = tester.create_openai_model(*config_and_inputs)
+        tester.check_openai_model_output(output_result)
+
+        output_result = tester.create_openai_double_heads(*config_and_inputs)
+        tester.check_openai_double_heads_output(output_result)
+        tester.check_openai_double_heads_loss_output(output_result)
+
+    @classmethod
+    def ids_tensor(cls, shape, vocab_size, rng=None, name=None):
+        """Creates a random int32 tensor of the shape within the vocab size."""
+        if rng is None:
+            rng = random.Random()
+
+        total_dims = 1
+        for dim in shape:
+            total_dims *= dim
+
+        values = []
+        for _ in range(total_dims):
+            values.append(rng.randint(0, vocab_size - 1))
+
+        return torch.tensor(data=values, dtype=torch.long).view(shape).contiguous()
+
+
+if __name__ == "__main__":
+    unittest.main()