Merge pull request #63 from chiragjn/opt_gpt2

Use `past` agrument for GPT2 to speed up decoding

nlpaug/augmenter/sentence/context_word_embs_sentence.py

@@ -12,7 +12,7 @@
 
 
 def init_context_word_embs_sentence_model(model_path, device, force_reload=False, temperature=1.0, top_k=None,
-                                          top_p=None):
+                                          top_p=None, return_past=True):
     global CONTEXT_WORD_EMBS_SENTENCE_MODELS
 
     model_name = os.path.basename(model_path)
@@ -23,16 +23,12 @@ def init_context_word_embs_sentence_model(model_path, device, force_reload=False
         return CONTEXT_WORD_EMBS_SENTENCE_MODELS[model_name]
 
     if 'xlnet' in model_path:
-        model = nml.XlNet(model_path, device=device, temperature=temperature, top_k=top_k, top_p=top_p)
+        model = nml.XlNet(model_path, device=device, temperature=temperature, top_k=top_k, top_p=top_p, return_past=return_past)
     elif 'gpt2' in model_path:
-        model = nml.Gpt2(model_path, device=device, temperature=temperature, top_k=top_k, top_p=top_p)
+        model = nml.Gpt2(model_path, device=device, temperature=temperature, top_k=top_k, top_p=top_p, return_past=return_past)
     else:
         raise ValueError('Model name value is unexpected. Only support XLNet and GPT2 model.')
 
-    CONTEXT_WORD_EMBS_SENTENCE_MODELS[model_name] = model
-    return model
-
-
 class ContextualWordEmbsForSentenceAug(SentenceAugmenter):
     # https://arxiv.org/pdf/1707.07328.pdf
     """
@@ -52,8 +48,8 @@ class ContextualWordEmbsForSentenceAug(SentenceAugmenter):
         Default value is False and suggesting to keep it as False if performance is the consideration.
     :param str name: Name of this augmenter
 
-    >>> import nlpaug.augmenter.word as naw
-    >>> aug = naw.ContextualWordEmbsForSentenceAug()
+    >>> import nlpaug.augmenter.sentence as nas
+    >>> aug = nas.ContextualWordEmbsForSentenceAug()
     """
 
     def __init__(self, model_path='xlnet-base-cased', temperature=1.0, top_k=100, top_p=None,
@@ -86,7 +82,8 @@ def insert(self, data):
         if data is None or data == '' or data.strip() == '':
             return data
 
-        max_try = 100
+        max_try = 30  # On average 30 should be enough to complete a sentence
+        past = None
         augmented_text = ''
 
         for _ in range(max_try):
@@ -95,7 +92,11 @@ def insert(self, data):
             if self.model_type in ['xlnet']:
                 text += ' ' + self.model.MASK_TOKEN
 
-            results = self.model.predict(text, n=1)
+            results = self.model.predict(text, n=1, past=past)
+
+            if self.model.return_past:
+                results, past = results
+
             new_word, proba = results[0]
 
             if new_word in self.SENTENCE_SEPARATOR:

nlpaug/model/lang_models/gpt2.py

@@ -12,7 +12,7 @@ class Gpt2(LanguageModels):
     # https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
     SUBWORD_PREFIX = 'Ġ'
 
-    def __init__(self, model_path='gpt2', temperature=1.0, top_k=None, top_p=None, device=None):
+    def __init__(self, model_path='gpt2', temperature=1.0, top_k=None, top_p=None, device=None, return_past=False):
         super().__init__(device, temperature=temperature, top_k=top_k, top_p=top_p)
         self.model_path = model_path
 
@@ -22,17 +22,21 @@ def __init__(self, model_path='gpt2', temperature=1.0, top_k=None, top_p=None, d
         self.model.to(self.device)
         self.model.eval()
 
+        self.return_past = return_past
+
     def id2token(self, _id):
         return self.tokenizer.decode(_id, clean_up_tokenization_spaces=True).strip()
 
-    def predict(self, text, target_word=None, n=1):
+    def predict(self, text, target_word=None, n=1, past=None):
         # Convert feature
         input_idxes = self.tokenizer.encode(text)
+        if past is not None:
+            input_idxes = input_idxes[-1:]
         input_idxes = torch.tensor(input_idxes, device=self.device).unsqueeze(0).repeat(1, 1)
 
         # Prediction
         with torch.no_grad():
-            outputs = self.model(input_idxes)
+            outputs = self.model(input_ids=input_idxes, past=past)
         target_token_logits = outputs[0][0][-1]  # GPT2 only predict last token
 
         # Selection
@@ -41,4 +45,8 @@ def predict(self, text, target_word=None, n=1):
         target_token_logits, target_token_idxes = self.filtering(target_token_logits, seed)
 
         results = self.pick(target_token_logits, target_word=target_word, n=n)
+        if self.return_past:
+            past = outputs[1]
+            results = (results, past,)
+
         return results

nlpaug/model/lang_models/xlnet.py

@@ -27,7 +27,7 @@ class XlNet(LanguageModels):
     NEW_PARAGRAPH_TOKEN = '<eop>'
 
     def __init__(self, model_path='xlnet-base-cased', temperature=1.0, top_k=None, top_p=None, padding_text=None,
-                 device=None):
+                 device=None, return_past=False):
         super().__init__(device, temperature=temperature, top_k=top_k, top_p=top_p)
         self.model_path = model_path
 
@@ -39,13 +39,16 @@ def __init__(self, model_path='xlnet-base-cased', temperature=1.0, top_k=None, t
         self.model.to(self.device)
         self.model.eval()
 
+        self.return_past = return_past
+
     def id2token(self, _id):
         return self.tokenizer.decode(_id, clean_up_tokenization_spaces=True).strip()
 
     def clean(self, text):
         return text.replace(self.NEW_PARAGRAPH_TOKEN, '').strip()
 
-    def predict(self, text, target_word=None, n=1):
+    def predict(self, text, target_word=None, n=1, past=None):
+        # xlnet does not support `past`, instead there is `mems` which works differently
         # Convert feature
         input_idxes = self.tokenizer.encode(text)
         concatenated_idxes = self.padding_text_idxes + input_idxes
@@ -72,4 +75,8 @@ def predict(self, text, target_word=None, n=1):
         target_token_logits, target_token_idxes = self.filtering(target_token_logits, seed)
 
         results = self.pick(target_token_logits, target_word=target_word, n=n)
-        return results
+
+        if self.return_past:
+            results = (results, past,)  # Only, for API compatibility, past is not used for xlnet
+
+        return results