Merge branch 'develop'

CHANGELOG.md

@@ -6,6 +6,20 @@ Most recent releases are shown at the top. Each release shows:
 - **Changed**: Additional parameters, changes to inputs or outputs, etc
 - **Fixed**: Bug fixes that don't change documented behaviour
 
+## 0.18.5 (2020-07-20)
+
+### New:
+- N/A
+
+### Changed
+- N/A
+
+### Fixed:
+- Changed `qa` module to use use 'Auto' when loading `QuestionAnswering` models and tokenizer
+- try `from_pt=True` for `qa` module if initial model-loading fails
+- use `get_hf_model_name` in `qa` module
+
+
 ## 0.18.4 (2020-07-17)
 
 ### New:

ktrain/text/qa/core.py

@@ -10,8 +10,10 @@
 from whoosh.qparser import QueryParser
 
 
-from transformers import TFBertForQuestionAnswering
-from transformers import BertTokenizer
+#from transformers import TFBertForQuestionAnswering
+#from transformers import BertTokenizer
+from transformers import TFAutoModelForQuestionAnswering
+from transformers import AutoTokenizer
 LOWCONF = -10000
 
 
@@ -23,8 +25,11 @@ class QA(ABC):
     def __init__(self, bert_squad_model='bert-large-uncased-whole-word-masking-finetuned-squad',
                  bert_emb_model='bert-base-uncased'):
         self.model_name = bert_squad_model
-        self.model = TFBertForQuestionAnswering.from_pretrained(self.model_name)
-        self.tokenizer = BertTokenizer.from_pretrained(self.model_name)
+        try:
+            self.model = TFAutoModelForQuestionAnswering.from_pretrained(self.model_name)
+        except:
+            self.model = TFAutoModelForQuestionAnswering.from_pretrained(self.model_name, from_pt=True)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
         self.maxlen = 512
         self.te = tpp.TransformerEmbedding(bert_emb_model, layers=[-2])
 
@@ -43,12 +48,19 @@ def predict_squad(self, document, question):
         assert len(segment_ids) == len(input_ids)
         n_ids = len(segment_ids)
         if n_ids < self.maxlen:
-            start_scores, end_scores = self.model(np.array([input_ids]), 
-                                             token_type_ids=np.array([segment_ids]))
+            input_ids = np.array([input_ids])
+            token_type_ids = np.array([segment_ids])
         else:
             #TODO: use different truncation strategies or run multiple inferences
-            start_scores, end_scores = self.model(np.array([input_ids[:self.maxlen]]), 
-                                             token_type_ids=np.array([segment_ids[:self.maxlen]]))
+            input_ids = np.array([input_ids[:self.maxlen]])
+            token_type_ids = np.array([segment_ids[:self.maxlen]])
+
+        # Added from: https://github.com/huggingface/transformers/commit/16ce15ed4bd0865d24a94aa839a44cf0f400ef50
+        if U.get_hf_model_name(self.model_name) in  ['xlm', 'roberta', 'distilbert']:
+            start_scores, end_scores = self.model(input_ids)
+        else:
+            start_scores, end_scores = self.model(input_ids, token_type_ids=token_type_ids)
+
         start_scores = start_scores[:,1:-1]
         end_scores = end_scores[:,1:-1]
         answer_start = np.argmax(start_scores)

ktrain/utils.py

@@ -492,3 +492,16 @@ def list2chunks(a, n):
     k, m = divmod(len(a), n)
     return (a[i * k + min(i, m):(i + 1) * k + min(i + 1, m)] for i in range(n))
 
+
+def get_hf_model_name(model_id):
+    parts = model_id.split('/')
+    if len(parts) == 1:
+        model_id = parts[0]
+    else:
+        model_id = '/'.join(parts[1:])
+    if model_id.startswith('xlm-roberta'): 
+        model_name = 'xlm-roberta'
+    else:
+        model_name = model_id.split('-')[0]
+    return model_name
+

ktrain/version.py

@@ -1,2 +1,2 @@
 __all__ = ['__version__']
-__version__ = '0.18.4'
+__version__ = '0.18.5'