cooelf/Retro-Reader
1
1import torch2from transformers import AlbertTokenizer, AlbertForSequenceClassification, AlbertForQuestionAnswering3import collections4import math5import gradio as gr6 7cls_modelPath = "./cls_model"8mrc_modelPath = "./model4"9 10tokenizer = AlbertTokenizer.from_pretrained(mrc_modelPath)11cls_model = AlbertForSequenceClassification.from_pretrained(cls_modelPath)12cls_model.eval()13mrc_model = AlbertForQuestionAnswering.from_pretrained(mrc_modelPath)14mrc_model.eval()15 16def _get_best_indexes(logits, n_best_size):17 """Get the n-best logits from a list."""18 index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)19 20 best_indexes = []21 for i in range(len(index_and_score)):22 if i >= n_best_size:23 break24 best_indexes.append(index_and_score[i][0])25 return best_indexes26 27def _compute_softmax(scores):28 """Compute softmax probability over raw logits."""29 if not scores:30 return []31 32 max_score = None33 for score in scores:34 if max_score is None or score > max_score:35 max_score = score36 37 exp_scores = []38 total_sum = 0.039 for score in scores:40 x = math.exp(score - max_score)41 exp_scores.append(x)42 total_sum += x43 44 probs = []45 for score in exp_scores:46 probs.append(score / total_sum)47 return probs48 49def get_qa_nbest(input_ids, start_logits, end_logits, seq_len, n_best_size=20, max_answer_length=30):50 score_null = 1000000 # large and positive51 prelim_predictions = []52 null_start_logit = 0 # the start logit at the slice with min null score53 null_end_logit = 0 # the end logit at the slice with min null score54 _PrelimPrediction = collections.namedtuple( # pylint: disable=invalid-name55 "PrelimPrediction",56 ["start_index", "end_index", "start_logit", "end_logit"])57 _NbestPrediction = collections.namedtuple( # pylint: disable=invalid-name58 "NbestPrediction", ["text", "start_logit", "end_logit"])59 60 start_indexes = _get_best_indexes(start_logits, n_best_size)61 end_indexes = _get_best_indexes(end_logits, n_best_size)62 63 feature_null_score = start_logits[0] + end_logits[0]64 if feature_null_score < score_null:65 score_null = feature_null_score66 for start_index in start_indexes:67 for end_index in end_indexes:68 if end_index < start_index:69 continue70 length = end_index - start_index + 171 if length > max_answer_length:72 continue73 if start_index >= seq_len:74 continue75 if end_index >= seq_len:76 continue77 prelim_predictions.append(78 _PrelimPrediction(79 start_index=start_index,80 end_index=end_index,81 start_logit=start_logits[start_index],82 end_logit=end_logits[end_index]))83 prelim_predictions = sorted(84 prelim_predictions,85 key=lambda x: (x.start_logit + x.start_logit),86 reverse=True)87 88 seen_predictions = {}89 nbest = []90 for pred in prelim_predictions:91 if len(nbest) >= n_best_size:92 break93 94 if pred.start_index > 0: # this is a non-null prediction\95 predict_answer_tokens = input_ids[0, pred.start_index: (pred.end_index + 1)]96 final_text = tokenizer.decode(predict_answer_tokens)97 if final_text in seen_predictions:98 continue99 seen_predictions[final_text] = True100 else:101 final_text = ""102 seen_predictions[final_text] = True103 104 105 nbest.append(106 _NbestPrediction(107 text=final_text,108 start_logit=pred.start_logit,109 end_logit=pred.end_logit))110 if "" not in seen_predictions:111 nbest.append(112 _NbestPrediction(113 text="",114 start_logit=null_start_logit,115 end_logit=null_end_logit))116 117 # In very rare edge cases we could only have single null prediction.118 # So we just create a nonce prediction in this case to avoid failure.119 if len(nbest) == 1:120 nbest.insert(0,121 _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))122 123 # In very rare edge cases we could have no valid predictions. So we124 # just create a nonce prediction in this case to avoid failure.125 if not nbest:126 nbest.append(127 _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))128 129 130 total_scores = []131 best_non_null_entry = None132 for entry in nbest:133 total_scores.append(entry.start_logit + entry.end_logit)134 if not best_non_null_entry:135 if entry.text:136 best_non_null_entry = entry137 138 probs = _compute_softmax(total_scores)139 nbest_json = []140 for (i, entry) in enumerate(nbest):141 output = collections.OrderedDict()142 output["text"] = entry.text143 output["probability"] = probs[i]144 output["start_logit"] = entry.start_logit145 output["end_logit"] = entry.end_logit146 nbest_json.append(output)147 148 score_diff = score_null - best_non_null_entry.start_logit - (149 best_non_null_entry.end_logit)150 151 return nbest_json, score_diff152 153def inference(context, question, reference):154 inputs = tokenizer(155 question,156 context,157 add_special_tokens=True,158 pad_to_max_length=True,159 max_length=512,160 return_tensors="pt"161 )162 163 seq_len = inputs.input_ids[0].tolist().index(0)164 165 with torch.no_grad():166 cls_outputs = cls_model(**inputs)167 qa_outputs = mrc_model(**inputs)168 169 cls_logits = cls_outputs.logits[0]170 cls_divide = cls_logits[1] - cls_logits[0]171 172 173 nbest, score_diff = get_qa_nbest(inputs.input_ids, qa_outputs.start_logits[0], qa_outputs.end_logits[0], seq_len=seq_len)174 175 thresh = -1.246073067188263176 177 print(cls_divide, score_diff)178 179 na_score = (0.5*cls_divide + 0.5*score_diff)*0.5180 if na_score > thresh:181 final_answer = f"<No Answer>. The question is not answerable according to the context."182 else:183 final_answer = nbest[0]["text"]184 return final_answer185 186demo = gr.Interface(187 fn=inference,188 inputs=[gr.inputs.Textbox(label="Context"), 189 gr.inputs.Textbox(label="Question"),190 gr.inputs.Textbox(label="Reference Answer (Optional)")],191 outputs=gr.outputs.Textbox(label="Output Answer"),192 examples = [193 ["The Norman dynasty had a major political, cultural and military impact on medieval Europe and even the Near East. The Normans were famed for their martial spirit and eventually for their Christian piety, becoming exponents of the Catholic orthodoxy into which they assimilated. They adopted the Gallo-Romance language of the Frankish land they settled, their dialect becoming known as Norman, Normaund or Norman French, an important literary language. The Duchy of Normandy, which they formed by treaty with the French crown, was a great fief of medieval France, and under Richard I of Normandy was forged into a cohesive and formidable principality in feudal tenure. The Normans are noted both for their culture, such as their unique Romanesque architecture and musical traditions, and for their significant military accomplishments and innovations. Norman adventurers founded the Kingdom of Sicily under Roger II after conquering southern Italy on the Saracens and Byzantines, and an expedition on behalf of their duke, William the Conqueror, led to the Norman conquest of England at the Battle of Hastings in 1066. Norman cultural and military influence spread from these new European centres to the Crusader states of the Near East, where their prince Bohemond I founded the Principality of Antioch in the Levant, to Scotland and Wales in Great Britain, to Ireland, and to the coasts of north Africa and the Canary Islands.", "Who was the duke in the battle of Hastings?", "William the Conqueror"],194 ["The Norman dynasty had a major political, cultural and military impact on medieval Europe and even the Near East. The Normans were famed for their martial spirit and eventually for their Christian piety, becoming exponents of the Catholic orthodoxy into which they assimilated. They adopted the Gallo-Romance language of the Frankish land they settled, their dialect becoming known as Norman, Normaund or Norman French, an important literary language. The Duchy of Normandy, which they formed by treaty with the French crown, was a great fief of medieval France, and under Richard I of Normandy was forged into a cohesive and formidable principality in feudal tenure. The Normans are noted both for their culture, such as their unique Romanesque architecture and musical traditions, and for their significant military accomplishments and innovations. Norman adventurers founded the Kingdom of Sicily under Roger II after conquering southern Italy on the Saracens and Byzantines, and an expedition on behalf of their duke, William the Conqueror, led to the Norman conquest of England at the Battle of Hastings in 1066. Norman cultural and military influence spread from these new European centres to the Crusader states of the Near East, where their prince Bohemond I founded the Principality of Antioch in the Levant, to Scotland and Wales in Great Britain, to Ireland, and to the coasts of north Africa and the Canary Islands.", "What type of major impact did the Norman dynasty have on modern Europe?", "<No Answer>"],195 ["Steam engines are external combustion engines, where the working fluid is separate from the combustion products. Non-combustion heat sources such as solar power, nuclear power or geothermal energy may be used. The ideal thermodynamic cycle used to analyze this process is called the Rankine cycle. In the cycle, water is heated and transforms into steam within a boiler operating at a high pressure. When expanded through pistons or turbines, mechanical work is done. The reduced-pressure steam is then condensed and pumped back into the boiler.", "What types of engines are steam engines?", "external combustion engines"],196 ["Steam engines are external combustion engines, where the working fluid is separate from the combustion products. Non-combustion heat sources such as solar power, nuclear power or geothermal energy may be used. The ideal thermodynamic cycle used to analyze this process is called the Rankine cycle. In the cycle, water is heated and transforms into steam within a boiler operating at a high pressure. When expanded through pistons or turbines, mechanical work is done. The reduced-pressure steam is then condensed and pumped back into the boiler.", "What ideal thermodynamic cycle analyzes the process by which solar engines work?", "<No Answer>"],197 ],198 title="Retrospective Reader for Machine Reading Comprehension",199 description=("<div style='text-align: center; margin: 0 auto;'>The model achieved the best performance at the SQuAD2.0 leaderboard. See more details at: <a href='https://aaai.org/papers/14506-retrospective-reader-for-machine-reading-comprehension/'>Paper</a> and <a href='https://github.com/cooelf/AwesomeMRC'>GitHub</a></div>"),200 )201 202demo.launch(debug=True)