-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_BiLSTM.py
296 lines (237 loc) · 11 KB
/
eval_BiLSTM.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
# Preprocessing
import numpy as np
import pandas as pd
from nltk.tokenize import word_tokenize
from sklearn.preprocessing import LabelEncoder
from keras.utils import to_categorical
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
import pickle
import sys
# Modeling
from keras.models import Model
from keras import backend as K
import os
# Evaluation
from keras.models import load_model
from sklearn.metrics import confusion_matrix
from utils import *
from sklearn.svm import OneClassSVM
from tqdm import tqdm
from sklearn.neighbors import LocalOutlierFactor
import pymysql.cursors
# GPU setting
dataset = sys.argv[1]
proportion = int(sys.argv[2])
logger = create_logger('BiLSTM_' + dataset)
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
if proportion==25:
gpu_id = "0"
elif proportion==50:
gpu_id = "2"
elif proportion==75:
gpu_id = "3"
set_allow_growth(gpu_id)
df, partition_to_n_row = load_single(dataset)
df['content_words'] = df['text'].apply(lambda s: word_tokenize(s))
df['words_len'] = df['content_words'].apply(lambda s: len(s))
texts = df['content_words'].tolist()
MAX_SEQ_LEN = None
MAX_NUM_WORDS = 10000
# filters without "," and "."
tokenizer = Tokenizer(num_words=MAX_NUM_WORDS, oov_token="<UNK>", filters='!"#$%&()*+-/:;<=>@[\]^_`{|}~')
tokenizer.fit_on_texts(texts)
word_index = tokenizer.word_index
sequences = tokenizer.texts_to_sequences(texts)
sequences_pad = pad_sequences(sequences, maxlen=MAX_SEQ_LEN, padding='post', truncating='post')
idx_train = (None, partition_to_n_row['train'])
idx_valid = (partition_to_n_row['train'], partition_to_n_row['train'] + partition_to_n_row['valid'])
idx_test = (partition_to_n_row['train'] + partition_to_n_row['valid'], None)
X_train = sequences_pad[idx_train[0]:idx_train[1]]
X_valid = sequences_pad[idx_valid[0]:idx_valid[1]]
X_test = sequences_pad[idx_test[0]:idx_test[1]]
df_train = df[idx_train[0]:idx_train[1]]
df_valid = df[idx_valid[0]:idx_valid[1]]
df_test = df[idx_test[0]:idx_test[1]]
y_train = df_train.label.reset_index(drop=True)
y_valid = df_valid.label.reset_index(drop=True)
y_test = df_test.label.reset_index(drop=True)
n_class = y_train.unique().shape[0]
n_class_seen = round(n_class * proportion/100)
for number in range(10):
print("start:", dataset, proportion, number)
with open('data/y_cols_' + dataset + "_" + str(proportion) + '_' + str(number) + '.pickle', 'rb') as handle:
d = pickle.load(handle)
y_cols_seen = d['y_cols_seen']
y_cols_unseen = d['y_cols_unseen']
print(y_cols_seen)
train_seen_idx = y_train[y_train.isin(y_cols_seen)].index
valid_seen_idx = y_valid[y_valid.isin(y_cols_seen)].index
X_train_seen = X_train[train_seen_idx]
y_train_seen = y_train[train_seen_idx]
X_valid_seen = X_valid[valid_seen_idx]
y_valid_seen = y_valid[valid_seen_idx]
le = LabelEncoder()
le.fit(y_train_seen)
y_train_idx = le.transform(y_train_seen)
y_valid_idx = le.transform(y_valid_seen)
y_train_onehot = to_categorical(y_train_idx)
y_valid_onehot = to_categorical(y_valid_idx)
y_test_mask = y_test.copy()
y_test_mask[y_test_mask.isin(y_cols_unseen)] = 'unseen'
train_data = (X_train_seen, y_train_onehot)
valid_data = (X_valid_seen, y_valid_onehot)
test_data = (X_test, y_test_mask)
# Load model
model = load_model('data/BiLSTM_' + dataset + "_" + str(proportion) + '_' + str(number) + '.h5')
y_pred_proba = model.predict(test_data[0])
y_pred_proba_train = model.predict(train_data[0])
classes = list(le.classes_) + ['unseen']
d_result = {
'all': defaultdict(dict),
'seen': defaultdict(dict),
'unseen': defaultdict(dict),
}
alpha = 2
method = "1Softmax (t=0.5)"
df_seen = pd.DataFrame(y_pred_proba, columns=le.classes_)
df_seen['unseen'] = 1 - df_seen.max(axis=1)
y_pred = df_seen.idxmax(axis=1)
cm = confusion_matrix(test_data[1], y_pred, classes)
f, d_result = get_score(cm, d_result, method)
method = "3DOC (Softmax)"
df_seen = pd.DataFrame(y_pred_proba, columns=le.classes_)
df_seen_train = pd.DataFrame(y_pred_proba_train, columns=le.classes_)
df_seen_train['y_true'] = y_train_seen.values
# Calcuate statistic threshold for unknown intent detection
col_to_threshold = {}
for col in y_cols_seen:
tmp = df_seen_train[df_seen_train['y_true']==col][[col, 'y_true']]
tmp = np.hstack([tmp[col], 2-tmp[col]])
threshold = tmp.mean() - alpha*tmp.std()
col_to_threshold[col] = threshold
col_to_threshold = {k: max([0.5, v])for k, v in col_to_threshold.items()}
masks = [df_seen[col]<threshold for col, threshold in col_to_threshold.items()]
is_reject = masks[0]
for mask in masks:
is_reject &= mask
df_seen['unseen'] = is_reject.astype(int)
y_pred = df_seen.idxmax(axis=1)
cm = confusion_matrix(test_data[1], y_pred, classes)
f, d_result = get_score(cm, d_result, method)
method = "4SofterMax"
get_logits = Model(inputs=model.input,
outputs=model.layers[-2].output)
get_pred = K.function([model.layers[-1].input],
[model.layers[-1].output])
# Find optimal temperature wrt logloss
logits_valid = get_logits.predict(valid_data[0])
logits = torch.from_numpy(logits_valid).float().cuda()
labels = torch.from_numpy(y_valid_idx).long().cuda()
modeT = ModelWithTemperature()
T, before_ece, after_ece = modeT.set_temperature(logits, labels)
T = max(1, T)
logits_test = get_logits.predict(test_data[0])
y_pred_proba_calibrated = get_pred([logits_test/T])[0]
logits_train = get_logits.predict(train_data[0])
y_pred_proba_train_calibrated = get_pred([logits_train/T])[0]
df_seen = pd.DataFrame(y_pred_proba_calibrated, columns=le.classes_)
df_seen_train = pd.DataFrame(y_pred_proba_train_calibrated, columns=le.classes_)
df_seen_train['y_true'] = y_train_seen.values
col_to_threshold = {}
for col in y_cols_seen:
tmp = df_seen_train[df_seen_train['y_true']==col][[col, 'y_true']]
tmp = np.hstack([tmp[col], 2-tmp[col]])
threshold = tmp.mean() - alpha*tmp.std()
col_to_threshold[col] = threshold
col_to_threshold = {k: max([0.5, v])for k, v in col_to_threshold.items()}
masks = [df_seen[col]<threshold for col, threshold in col_to_threshold.items()]
is_reject_TS = masks[0]
for mask in masks:
is_reject_TS &= mask
df_seen['unseen'] = is_reject_TS.astype(int)
y_pred = df_seen.idxmax(axis=1)
cm = confusion_matrix(test_data[1], y_pred, classes)
f, d_result = get_score(cm, d_result, method)
method = "5LOF"
get_deep_feature = Model(inputs=model.input,
outputs=model.layers[-3].output)
feature_test = get_deep_feature.predict(test_data[0])
path_lof = 'data/lof_' + dataset + "_" + str(proportion) + '_' + str(number) + '.pickle'
try:
lof = pickle.load(open(path_lof, "rb"))
print("pretrain LOF found:", path_lof)
except (OSError, IOError) as e:
feature_train = get_deep_feature.predict(train_data[0])
lof = LocalOutlierFactor(n_neighbors=20, contamination=0.05, novelty=True, n_jobs=-1)
lof.fit(feature_train)
pickle.dump(lof, open(path_lof, "wb"))
y_pred_lof = pd.Series(lof.predict(feature_test))
df_seen = pd.DataFrame(y_pred_proba, columns=le.classes_)
df_seen['unseen'] = 0
y_pred = df_seen.idxmax(axis=1)
y_pred[y_pred_lof[y_pred_lof==-1].index]='unseen'
cm = confusion_matrix(test_data[1], y_pred, classes)
f, d_result = get_score(cm, d_result, method)
#### Transform SofterMax score into probability through Platt Scaling (Pseudo code)
# centralize probability (m sample, n classes) = calibrated_probability(m, n) - probability threshold(1, n)
# novelty score (m, 1) = max(centralize probability)
# novelty probability (m, 1) = Platt Scaling(novelty score)
df_seen = pd.DataFrame(y_pred_proba_calibrated, columns=le.classes_).copy()
for col, threshold in col_to_threshold.items():
df_seen[col] = df_seen[col]-threshold
decision_function = df_seen.max(axis=1)
predict = (decision_function>0).astype(int) # 1=inliner, 0=outlier
decision_function = np.array(decision_function).reshape(-1,1)
# Standardization (novelty score)
ss = StandardScaler()
decision_function_z = ss.fit_transform(decision_function)
# Platt scaling (transform score into probability)
lr = LogisticRegression('l1', solver='liblinear', C=1, class_weight='balanced', max_iter=1000)
lr.fit(decision_function_z, predict)
predict_prob_sm =lr.predict_proba(decision_function_z)[:, 0]
#### Transform LOF score into probability through Platt Scaling
path_lof = 'data/lof_' + dataset + "_" + str(proportion) + '_' + str(number) + '.pickle'
try:
lof = pickle.load(open(path_lof, "rb"))
print("pretrain LOF found:", path_lof)
except (OSError, IOError) as e:
feature_train = get_deep_feature.predict(train_data[0])
lof = LocalOutlierFactor(n_neighbors=20, contamination=0.05, novelty=True, n_jobs=-1)
lof.fit(feature_train)
pickle.dump(lof, open(path_lof, "wb"))
# outlier score threshold
score_samples = lof.score_samples(feature_test)
factor_ = lof.negative_outlier_factor_
decision_function = score_samples - lof.offset_
predict = (decision_function>0).astype(int)
# Calibrate discrete prediction{1, 0} into probability(1~0)
ss = StandardScaler()
decision_function = np.reshape(decision_function, (-1, 1))
decision_function_z = ss.fit_transform(decision_function)
lr = LogisticRegression('l1', solver='liblinear', C=1, class_weight='balanced', max_iter=1000)
lr.fit(decision_function_z, predict)
predict_prob_lof =lr.predict_proba(decision_function_z)[:, 0]
method="6SMDN"
df_SMDN = pd.DataFrame([predict_prob_lof, predict_prob_sm]).T.copy()
df_SMDN.columns = ['LOF', 'SofterMax']
df_SMDN['unseen'] = df_SMDN.mean(axis=1)
df_seen = pd.DataFrame(y_pred_proba_calibrated, columns=le.classes_).copy()
df_seen['unseen'] = (df_SMDN['unseen']>0.5).astype(int)
y_pred = df_seen.idxmax(axis=1)
cm = confusion_matrix(test_data[1], y_pred, classes)
f, d_result = get_score(cm, d_result, method)
# Save the result
results = []
for part, d in d_result.items():
for method, score in d.items():
results.append([dataset, proportion, number, part, method, float(score)])
connection = pymysql.connect(host='localhost', user='root', password='', db='KBS',
charset='utf8mb4', cursorclass=pymysql.cursors.DictCursor)
with connection.cursor() as cursor:
# Create a new record
sql = "INSERT INTO `result` (`dataset`, `proportion`, `number`, `part`, `method`, `score`, `temperature`, `before_ece`, `after_ece`) VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s)"
for result in results:
cursor.execute(sql, result+[T, before_ece, after_ece])
connection.commit()
connection.close()