File size: 13,781 Bytes
91cb36e 4c48e8e 91cb36e 4c48e8e 91cb36e 25ca7c8 a343bb9 25ca7c8 0da9196 91cb36e 4c48e8e 91cb36e 0da9196 91cb36e 4c48e8e 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 91cb36e a343bb9 0da9196 a343bb9 0da9196 4c48e8e da14a77 4c48e8e da14a77 4c48e8e 212d174 da14a77 4c48e8e da14a77 212d174 4c48e8e 212d174 4c48e8e 91cb36e 3a7acca 91cb36e a343bb9 91cb36e 0da9196 91cb36e 0da9196 a343bb9 91cb36e 0da9196 91cb36e 0da9196 91cb36e 0da9196 4c48e8e a343bb9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 |
from typing import Dict, List, Any
from scipy.special import softmax
import numpy as np
import weakref
import re
from utils import clean_str, clean_str_nopunct
import torch
from utils import MultiHeadModel, BertInputBuilder, get_num_words, MATH_PREFIXES, MATH_WORDS
import transformers
from transformers import BertTokenizer, BertForSequenceClassification
transformers.logging.set_verbosity_debug()
UPTAKE_MODEL = 'ddemszky/uptake-model'
REASONING_MODEL = 'ddemszky/student-reasoning'
QUESTION_MODEL = 'ddemszky/question-detection'
FOCUSING_QUESTION_MODEL = 'ddemszky/focusing-questions'
class Utterance:
def __init__(self, speaker, text, uid=None,
transcript=None, starttime=None, endtime=None, **kwargs):
self.speaker = speaker
self.text = text
self.uid = uid
self.starttime = starttime
self.endtime = endtime
self.transcript = weakref.ref(transcript) if transcript else None
self.props = kwargs
self.num_math_terms = None
self.math_terms = None
self.uptake = None
self.reasoning = None
self.question = None
self.focusing_question = None
def get_clean_text(self, remove_punct=False):
if remove_punct:
return clean_str_nopunct(self.text)
return clean_str(self.text)
def get_num_words(self):
return get_num_words(self.text)
def to_dict(self):
return {
'speaker': self.speaker,
'text': self.text,
'uid': self.uid,
'starttime': self.starttime,
'endtime': self.endtime,
'uptake': self.uptake,
'reasoning': self.reasoning,
'question': self.question,
'focusingQuestion': self.focusing_question,
'numMathTerms': self.num_math_terms,
'mathTerms': self.math_terms,
**self.props
}
def __repr__(self):
return f"Utterance(speaker='{self.speaker}'," \
f"text='{self.text}', uid={self.uid}," \
f"starttime={self.starttime}, endtime={self.endtime}, props={self.props})"
class Transcript:
def __init__(self, **kwargs):
self.utterances = []
self.params = kwargs
def add_utterance(self, utterance):
utterance.transcript = weakref.ref(self)
self.utterances.append(utterance)
def get_idx(self, idx):
if idx >= len(self.utterances):
return None
return self.utterances[idx]
def get_uid(self, uid):
for utt in self.utterances:
if utt.uid == uid:
return utt
return None
def length(self):
return len(self.utterances)
def to_dict(self):
return {
'utterances': [utterance.to_dict() for utterance in self.utterances],
**self.params
}
def __repr__(self):
return f"Transcript(utterances={self.utterances}, custom_params={self.params})"
class QuestionModel:
def __init__(self, device, tokenizer, input_builder, max_length=300, path=QUESTION_MODEL):
print("Loading models...")
self.device = device
self.tokenizer = tokenizer
self.input_builder = input_builder
self.max_length = max_length
self.model = MultiHeadModel.from_pretrained(
path, head2size={"is_question": 2})
self.model.to(self.device)
def run_inference(self, transcript):
self.model.eval()
with torch.no_grad():
for i, utt in enumerate(transcript.utterances):
if "?" in utt.text:
utt.question = 1
else:
text = utt.get_clean_text(remove_punct=True)
instance = self.input_builder.build_inputs([], text,
max_length=self.max_length,
input_str=True)
output = self.get_prediction(instance)
print(output)
utt.question = np.argmax(
output["is_question_logits"][0].tolist())
def get_prediction(self, instance):
instance["attention_mask"] = [[1] * len(instance["input_ids"])]
for key in ["input_ids", "token_type_ids", "attention_mask"]:
instance[key] = torch.tensor(
instance[key]).unsqueeze(0) # Batch size = 1
instance[key].to(self.device)
output = self.model(input_ids=instance["input_ids"],
attention_mask=instance["attention_mask"],
token_type_ids=instance["token_type_ids"],
return_pooler_output=False)
return output
class ReasoningModel:
def __init__(self, device, tokenizer, input_builder, max_length=128, path=REASONING_MODEL):
print("Loading models...")
self.device = device
self.tokenizer = tokenizer
self.input_builder = input_builder
self.max_length = max_length
self.model = BertForSequenceClassification.from_pretrained(path)
self.model.to(self.device)
def run_inference(self, transcript, min_num_words=8):
self.model.eval()
with torch.no_grad():
for i, utt in enumerate(transcript.utterances):
if utt.get_num_words() >= min_num_words:
instance = self.input_builder.build_inputs([], utt.text,
max_length=self.max_length,
input_str=True)
output = self.get_prediction(instance)
utt.reasoning = np.argmax(output["logits"][0].tolist())
def get_prediction(self, instance):
instance["attention_mask"] = [[1] * len(instance["input_ids"])]
for key in ["input_ids", "token_type_ids", "attention_mask"]:
instance[key] = torch.tensor(
instance[key]).unsqueeze(0) # Batch size = 1
instance[key].to(self.device)
output = self.model(input_ids=instance["input_ids"],
attention_mask=instance["attention_mask"],
token_type_ids=instance["token_type_ids"])
return output
class UptakeModel:
def __init__(self, device, tokenizer, input_builder, max_length=120, path=UPTAKE_MODEL):
print("Loading models...")
self.device = device
self.tokenizer = tokenizer
self.input_builder = input_builder
self.max_length = max_length
self.model = MultiHeadModel.from_pretrained(path, head2size={"nsp": 2})
self.model.to(self.device)
def run_inference(self, transcript, min_prev_words, uptake_speaker=None):
self.model.eval()
prev_num_words = 0
prev_utt = None
with torch.no_grad():
for i, utt in enumerate(transcript.utterances):
if ((uptake_speaker is None) or (utt.speaker == uptake_speaker)) and (prev_num_words >= min_prev_words):
textA = prev_utt.get_clean_text(remove_punct=False)
textB = utt.get_clean_text(remove_punct=False)
instance = self.input_builder.build_inputs([textA], textB,
max_length=self.max_length,
input_str=True)
output = self.get_prediction(instance)
utt.uptake = int(
softmax(output["nsp_logits"][0].tolist())[1] > .8)
prev_num_words = utt.get_num_words()
prev_utt = utt
def get_prediction(self, instance):
instance["attention_mask"] = [[1] * len(instance["input_ids"])]
for key in ["input_ids", "token_type_ids", "attention_mask"]:
instance[key] = torch.tensor(
instance[key]).unsqueeze(0) # Batch size = 1
instance[key].to(self.device)
output = self.model(input_ids=instance["input_ids"],
attention_mask=instance["attention_mask"],
token_type_ids=instance["token_type_ids"],
return_pooler_output=False)
return output
class FocusingQuestionModel:
def __init__(self, device, tokenizer, input_builder, max_length=128, path=FOCUSING_QUESTION_MODEL):
print("Loading models...")
self.device = device
self.tokenizer = tokenizer
self.input_builder = input_builder
self.model = BertForSequenceClassification.from_pretrained(path)
self.model.to(self.device)
self.max_length = max_length
def run_inference(self, transcript, min_focusing_words=0, uptake_speaker=None):
self.model.eval()
with torch.no_grad():
for i, utt in enumerate(transcript.utterances):
if utt.speaker != uptake_speaker or uptake_speaker is None:
utt.focusing_question = None
continue
if utt.get_num_words() < min_focusing_words:
utt.focusing_question = None
continue
instance = self.input_builder.build_inputs([], utt.text, max_length=self.max_length, input_str=True)
output = self.get_prediction(instance)
utt.focusing_question = np.argmax(output["logits"][0].tolist())
def get_prediction(self, instance):
instance["attention_mask"] = [[1] * len(instance["input_ids"])]
for key in ["input_ids", "token_type_ids", "attention_mask"]:
instance[key] = torch.tensor(
instance[key]).unsqueeze(0) # Batch size = 1
instance[key].to(self.device)
output = self.model(input_ids=instance["input_ids"],
attention_mask=instance["attention_mask"],
token_type_ids=instance["token_type_ids"])
return output
def load_math_terms():
math_terms = []
math_terms_dict = {}
for term in MATH_WORDS:
if term in MATH_PREFIXES:
math_terms_dict[f"(^|[^a-zA-Z]){term}(s|es)?([^a-zA-Z]|$)"] = term
math_terms.append(f"(^|[^a-zA-Z]){term}(s|es)?([^a-zA-Z]|$)")
else:
math_terms.append(term)
math_terms_dict[term] = term
return math_terms, math_terms_dict
def run_math_density(transcript):
math_terms, math_terms_dict = load_math_terms()
sorted_terms = sorted(math_terms, key=len, reverse=True)
for i, utt in enumerate(transcript.utterances):
text = utt.get_clean_text(remove_punct=False)
num_matches = 0
matched_positions = set()
match_list = []
for term in sorted_terms:
matches = list(re.finditer(term, text, re.IGNORECASE))
# Filter out matches that share positions with longer terms
matches = [match for match in matches if not any(match.start() in range(existing[0], existing[1]) for existing in matched_positions)]
if len(matches) > 0:
match_list.append(math_terms_dict[term])
# Update existing match positions
matched_positions.update((match.start(), match.end()) for match in matches)
num_matches += len(matches)
utt.num_math_terms = num_matches
utt.math_terms = match_list
class EndpointHandler():
def __init__(self, path="."):
print("Loading models...")
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
self.input_builder = BertInputBuilder(tokenizer=self.tokenizer)
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
"""
data args:
inputs (:obj: `list`):
List of dicts, where each dict represents an utterance; each utterance object must have a `speaker`,
`text` and `uid`and can include list of custom properties
parameters (:obj: `dict`)
Return:
A :obj:`list` | `dict`: will be serialized and returned
"""
# get inputs
utterances = data.pop("inputs", data)
params = data.pop("parameters", None)
print("EXAMPLES")
for utt in utterances[:3]:
print("speaker %s: %s" % (utt["speaker"], utt["text"]))
transcript = Transcript(filename=params.pop("filename", None))
for utt in utterances:
transcript.add_utterance(Utterance(**utt))
print("Running inference on %d examples..." % transcript.length())
uptake_speaker = params.pop("uptake_speaker", None)
# Uptake
uptake_model = UptakeModel(
self.device, self.tokenizer, self.input_builder)
uptake_model.run_inference(transcript, min_prev_words=params['uptake_min_num_words'],
uptake_speaker=uptake_speaker)
# Reasoning
reasoning_model = ReasoningModel(
self.device, self.tokenizer, self.input_builder)
reasoning_model.run_inference(transcript)
# Question
question_model = QuestionModel(
self.device, self.tokenizer, self.input_builder)
question_model.run_inference(transcript)
# Focusing Question
focusing_question_model = FocusingQuestionModel(
self.device, self.tokenizer, self.input_builder)
focusing_question_model.run_inference(transcript, uptake_speaker=uptake_speaker)
run_math_density(transcript)
return transcript.to_dict() |