-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathmodel_evaluation.py
386 lines (314 loc) · 17.4 KB
/
model_evaluation.py
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
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
"""Contains wrappers for different evaluation functions to be used primarily as
reward functions for RL model"""
import numpy as np
import os
import config
import utils
import data
from train_ESIM import ESIM_pred, load_ESIM_model
import encoder_models
#%%
from eval_metrics.bleu.bleu_scorer import BleuScorer
from eval_metrics.rouge.rouge import Rouge
from eval_metrics.cider.cider_scorer import CiderScorer
from eval_metrics.meteor.meteor import Meteor
# Define conventional automatic metrics warppers from eval_metrics package
def BLEU_score(target_sentence, pred_sentence, n_tokens=1):
"""Returns BLEU score at specified n-gram level for a given target and predicted sentence pair"""
try:
# Set n to BLEU score level
bleu_scorer = BleuScorer(n=n_tokens)
bleu_scorer += (pred_sentence[0], target_sentence)
BLEU_score, _ = bleu_scorer.compute_score()
return np.around(BLEU_score[n_tokens-1], 4)
except:
print('rejected sentence: ', pred_sentence)
def ROUGE_score(target_sentence, pred_sentence):
"""Returns ROUGE score for a given target and predicted sentence pair"""
try:
rouge = Rouge()
ROUGE_score = rouge.calc_score(pred_sentence, target_sentence)
return np.around(ROUGE_score, 4)
except:
print('rejected sentence: ', pred_sentence)
def CIDER_score(target_sentence, pred_sentence):
"""Returns CIDER score for a given target and predicted sentence pair"""
try:
cider_scorer = CiderScorer(idf_terms_path=os.path.join(config.PATH, 'eval_metrics/cider/idf_terms'))
cider_scorer += (pred_sentence[0], target_sentence)
CIDER_score, _ = cider_scorer.compute_score()
return np.around(CIDER_score, 4)
except:
print('rejected sentence: ', pred_sentence)
def METEOR_score(target_sentence, pred_sentence):
"""Returns METEOR score for a given target and predicted sentence pair"""
try:
meteor = Meteor()
METEOR_score, _ = meteor.compute_score(pred_sentence,
target_sentence)
return np.around(METEOR_score, 4)
except:
print("Java did not execute properly.")
#%%
def tokenize(input_sentence):
"""Converts an input sentence to a set of tokens after applying preprocessing"""
preprocessed_sentence = data.preprocess(input_sentence, remove_punct=True, lower_case=True)
tokens = preprocessed_sentence.split()
return tokens
def sentence_similarity(target, pred, similarity_model):
"""Calculates the cosine similarity between the sentence embeddings of a target and predicted pair
using the embedding model specified"""
try:
cosine_sim = utils.cosine_similarity(similarity_model.sentence_embedding(target).view(1,-1),
similarity_model.sentence_embedding(pred).view(1,-1))
return np.around(cosine_sim.item(),4)
except:
print('similarity rejected sentence: ', pred)
return 0.01
def sentence_length(input_sentence, min_value=6, max_value=12):
"""Returns the sentence length score for use as an auxiliary reward function"""
input_value = np.clip(len(input_sentence.split()), min_value, max_value)
return 1 - ((input_value - min_value) / (max_value - min_value))
def avg_word_frequency(input_sentence, vocab_index, total_word_count):
"""Returns the average word frequency for use as an auxiliary reward function"""
return np.mean([vocab_index.word2count[word] for word in input_sentence.split()]) / total_word_count
def rare_word_prop(input_sentence, vocab_index, rare_thresh=10):
"""Returns the proportion of rare words in a sentence for use as an auxiliary reward function"""
try:
assert len(input_sentence) > 0, 'Sentence is empty'
input_words = input_sentence.split()
n_rare_words = 0
for word in input_words:
if vocab_index.word2count[word] <= rare_thresh:
n_rare_words += 1
else:
pass
return n_rare_words / len(input_words)
except:
print('rejected sentence: ', input_sentence)
return 0.01
def word_syllable_count(input_word):
"""Returns the syllable count for a word"""
"""source: https://stackoverflow.com/questions/46759492/syllable-count-in-python"""
vowels = "aeiouyAEIOUY"
count = 0
prior_letter = None
try:
for idx, letter in enumerate(input_word):
if (idx == 0) and (letter in vowels):
count += 1
elif (letter in vowels) and (prior_letter not in vowels):
count += 1
prior_letter = letter
if input_word.endswith("e"):
count -= 1
return max(1, count)
except:
print('rejected word: ', input_word)
return 1
def scaled_sent_syllable_count(input_sentence, max_avg_n_syllables=2):
"""Returns the average syllable count for a sentence for use as an auxiliary reward function"""
try:
assert len(input_sentence) > 0, 'Sentence is empty'
avg_syllable_count = np.mean([word_syllable_count(word) for word in input_sentence.split()])
return np.min([(avg_syllable_count / max_avg_n_syllables), 1])
except:
print('rejected sentence: ', input_sentence)
return 0.01
def libertarian_pred(input_sentence, BERT_model, logr_model, std_scaler):
"""Returns the probability a given sentence is a comment from a Libertarian subreddit rather than
an Anarchist / Socialist subreddit based on a trained model for use as an auxiliary reward function"""
try:
encoded_sent = BERT_model.sentence_embedding(input_sentence).reshape(1, -1)
standardized_sent = std_scaler.transform(encoded_sent)
probs = logr_model.predict_proba(standardized_sent)
return probs[0][1]
except:
print('rejected political sentence: ', input_sentence)
return 0.01
def performance_metrics(target_sentence, pred_sentence, similarity_model=None, fluency_model=None,
ESIM_model=None, logr_model=None, std_scaler=None, similarity_dist=None, fluency_dist=None,
ESIM_dist=None, vocab_index=None, metric='BLEU1'):
"""The main pipeline which handles applying the appropriate reward function and handles
the relevant models / data inputs required"""
if metric == 'BLEU1':
return BLEU_score(target_sentence=[target_sentence], pred_sentence=[pred_sentence], n_tokens=1)
if metric == 'BLEU2':
return BLEU_score(target_sentence=[target_sentence], pred_sentence=[pred_sentence], n_tokens=2)
elif metric == 'ROUGE':
return ROUGE_score(target_sentence=[target_sentence], pred_sentence=[pred_sentence])
elif metric == 'CIDER':
return CIDER_score(target_sentence=[target_sentence], pred_sentence=[pred_sentence])
elif metric == 'METEOR':
return METEOR_score(target_sentence=[target_sentence], pred_sentence=[pred_sentence])
elif metric == 'FLUENCY':
return fluency_model.fluency_score(pred_sentence)
elif metric == 'PARA':
return sentence_similarity(target_sentence, pred_sentence, similarity_model)
elif metric == 'PARA_F':
similarity_score = sentence_similarity(target_sentence, pred_sentence, similarity_model)
fluency_score = fluency_model.fluency_score(pred_sentence)
scaled_similarity_score = utils.cdf_score(similarity_dist, similarity_score)
scaled_fluency_score = utils.cdf_score(fluency_dist, fluency_score)
return np.mean([scaled_similarity_score, scaled_fluency_score])
elif metric == 'PARASIM':
similarity_score = sentence_similarity(target_sentence, pred_sentence, similarity_model)
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
scaled_similarity_score = utils.cdf_score(similarity_dist, similarity_score)
scaled_ESIM_score = utils.cdf_score(ESIM_dist, ESIM_score)
return np.mean([scaled_similarity_score, scaled_ESIM_score])
elif metric == 'PARASIM_F':
similarity_score = sentence_similarity(target_sentence, pred_sentence, similarity_model)
fluency_score = fluency_model.fluency_score(pred_sentence)
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
scaled_similarity_score = utils.cdf_score(similarity_dist, similarity_score)
scaled_fluency_score = utils.cdf_score(fluency_dist, fluency_score)
scaled_ESIM_score = utils.cdf_score(ESIM_dist, ESIM_score)
return np.mean([scaled_similarity_score, scaled_fluency_score, scaled_ESIM_score])
elif metric == 'ESIM':
return ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
elif metric == 'ESIM_short':
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
length_score = sentence_length(pred_sentence)
return 0.6 * ESIM_score + 0.4 * length_score
elif metric == 'ESIM_syllables':
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
syllable_score = scaled_sent_syllable_count(pred_sentence, max_avg_n_syllables=2)
return 0.6 * ESIM_score + 0.4 * syllable_score
elif metric == 'ESIM_rare':
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
rare_score = rare_word_prop(pred_sentence, vocab_index, rare_thresh=2500)
return 0.6 * ESIM_score + 0.4 * rare_score
elif metric == 'ESIM_F':
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
fluency_score = fluency_model.fluency_score(pred_sentence)
scaled_ESIM_score = utils.cdf_score(ESIM_dist, ESIM_score)
scaled_fluency_score = utils.cdf_score(fluency_dist, fluency_score)
return np.mean([scaled_ESIM_score, scaled_fluency_score])
elif metric == 'ESIM_libertarian':
ESIM_score = ESIM_pred([[target_sentence, pred_sentence]], ESIM_model, temperature=2).item()
libertarian_score = libertarian_pred(pred_sentence, similarity_model, logr_model, std_scaler)
return 0.6 * ESIM_score + 0.4 * libertarian_score
def init_eval_models(reward_function='BLEU1', similarity_model_name='BERT', ESIM_model_name='ESIM_noisy_3'):
"""Initializes the appropriate models / data for use in the performance metrics evaluation.
The following fields are being initialized:
similarity_model, fluency_model, ESIM_model, \
logr_model, std_scaler, \
similarity_dist, fluency_dist, ESIM_dist"""
if reward_function == 'FLUENCY':
return None, encoder_models.GPTLanguageModel(), None, \
None, None, \
None, None, None
if reward_function == 'PARA':
if similarity_model_name =='BERT':
return encoder_models.BERTEncoder(), None, None, \
None, None, \
None, None, None
elif similarity_model_name =='InferSent':
return encoder_models.InferSentEncoder(), None, None, \
None, None, \
None, None, None
elif reward_function == 'PARA_F':
if similarity_model_name =='BERT':
similarity_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'BERT_dist.npy'))
fluency_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'fluency_dist.npy'))
return encoder_models.BERTEncoder(), encoder_models.GPTLanguageModel(), None, \
None, None, \
similarity_dist, fluency_dist, None
elif similarity_model_name =='InferSent':
similarity_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'InferSent_dist.npy'))
fluency_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'fluency_dist.npy'))
return encoder_models.InferSentEncoder(), encoder_models.GPTLanguageModel(), None, \
None, None, \
similarity_dist, fluency_dist, None
elif reward_function == 'PARASIM':
if similarity_model_name =='BERT':
similarity_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'BERT_dist.npy'))
ESIM_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'ESIM_dist.npy'))
return encoder_models.BERTEncoder(), None, load_ESIM_model(ESIM_model_name), \
None, None, \
similarity_dist, None, ESIM_dist
elif similarity_model_name =='InferSent':
similarity_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'InferSent_dist.npy'))
ESIM_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'ESIM_dist.npy'))
return encoder_models.InferSentEncoder(), None, load_ESIM_model(ESIM_model_name), \
None, None, \
similarity_dist, None, ESIM_dist
elif reward_function == 'PARASIM_F':
if similarity_model_name =='BERT':
similarity_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'BERT_dist.npy'))
fluency_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'fluency_dist.npy'))
ESIM_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'ESIM_dist.npy'))
return encoder_models.BERTEncoder(), encoder_models.GPTLanguageModel(), load_ESIM_model(ESIM_model_name), \
None, None, \
similarity_dist, fluency_dist, ESIM_dist
elif similarity_model_name =='InferSent':
similarity_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'InferSent_dist.npy'))
fluency_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'fluency_dist.npy'))
ESIM_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'ESIM_dist.npy'))
return encoder_models.InferSentEncoder(), encoder_models.GPTLanguageModel(), load_ESIM_model(ESIM_model_name), \
None, None, \
similarity_dist, fluency_dist, ESIM_dist
elif reward_function == 'ESIM':
return None, None, load_ESIM_model(ESIM_model_name), \
None, None, \
None, None, None
elif reward_function == 'ESIM_short':
return None, None, load_ESIM_model(ESIM_model_name), \
None, None, \
None, None, None
elif reward_function == 'ESIM_rare':
return None, None, load_ESIM_model(ESIM_model_name), \
None, None, \
None, None, None
elif reward_function == 'ESIM_F':
fluency_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'fluency_dist.npy'))
ESIM_dist = data.load_np_data(os.path.join(config.saved_SM_dist_path, 'ESIM_dist.npy'))
return None, encoder_models.GPTLanguageModel(), load_ESIM_model(ESIM_model_name), \
None, None, \
None, fluency_dist, ESIM_dist
elif reward_function == 'ESIM_syllables':
return None, None, load_ESIM_model(ESIM_model_name), \
None, None, \
None, None, None
elif reward_function == 'ESIM_libertarian':
logr_model = data.load_vocab_index(os.path.join(
config.saved_reddit_model_path, 'anarsoc_libertar_logr.pickle'))
std_scaler = data.load_vocab_index(os.path.join(
config.saved_reddit_model_path, 'std_scaler.pickle'))
return encoder_models.BERTEncoder(), None, load_ESIM_model(ESIM_model_name), \
logr_model, std_scaler, \
None, None, None
else:
return None, None, None, None, None, None, None, None
#%% ------------------------------- ARCHIVE -------------------------------
# Prior defined versions of BLEU and ROUGE
#from collections import Counter
#from rouge.rouge import rouge_n_sentence_level
#from nltk import bigrams
#
#def BLEU_score(target_tokens, pred_tokens, ngram = 'unigram'):
# if ngram == 'bigram':
# target_tokens = list(bigrams(target_tokens))
# pred_tokens = list(bigrams(pred_tokens))
#
# word_counts = Counter(target_tokens)
# score = 0
#
# for token in pred_tokens:
# if word_counts[token] > 0:
# word_counts[token] -=1
# score += 1
# score /= len(target_tokens)
# return score
#
#def ROUGE_score(target, pred, ngram = 'unigram'):
# """ Note: rouge n_sentence_level has hypothesis and reference positions swapped"""
# if ngram == 'unigram':
# _, _, score = rouge_n_sentence_level(pred, target, 1)
# return score
# if ngram == 'bigram':
# _, _, score = rouge_n_sentence_level(pred, target, 2)
# return score
# else:
# print("Please select either: 'unigram' or 'bigram'")