-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_BiLSTM.py
147 lines (119 loc) · 5.28 KB
/
train_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
# Preprocessing
import os
import sys
import pickle
from sklearn.preprocessing import LabelEncoder
from keras.utils import to_categorical
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
# Modeling
from keras.callbacks import ModelCheckpoint, EarlyStopping
from utils import *
from models import BiLSTM
# GPU setting
dataset = sys.argv[1]
proportion = int(sys.argv[2])
if proportion==25:
gpu_id = "0"
elif proportion==50:
gpu_id = "1"
elif proportion==75:
gpu_id = "3"
logger = create_logger('BiLSTM_' + dataset)
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
set_allow_growth(gpu_id)
df, partition_to_n_row = load_single(dataset)
# Preprocessing
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)
# Load embedding
path = '/data/disk1/tony/'
EMBEDDING_FILE = path + 'glove.6B.300d.txt'
EMBEDDING_DIM = 300
MAX_SEQ_LEN = None # BiLSTM
# MAX_SEQ_LEN = df.words_len.max() # TextCNN
MAX_NB_WORDS = 10000
MAX_FEATURES = min(MAX_NB_WORDS, len(word_index)) + 1
def get_coefs(word,*arr): return word, np.asarray(arr, dtype='float32')
embeddings_index = dict(get_coefs(*o.strip().split()) for o in open(EMBEDDING_FILE))
all_embs = np.stack(embeddings_index.values())
emb_mean, emb_std = all_embs.mean(), all_embs.std()
# embedding_matrix的长度多一行，不存在embedding的词的值都为0 (pad)
embedding_matrix = np.random.normal(emb_mean, emb_std, (MAX_FEATURES, EMBEDDING_DIM))
for word, i in word_index.items():
if i >= MAX_FEATURES: continue
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None: embedding_matrix[i] = embedding_vector
# Repeat experiments for ten times
for number in range(10):
n_class = y_train.unique().shape[0]
n_class_seen = round(n_class * proportion/100)
# Randomly choose seen class
y_cols = y_train.unique()
y_vc = y_train.value_counts()
y_vc = y_vc / y_vc.sum()
y_cols_seen = np.random.choice(y_vc.index, n_class_seen, p=y_vc.values, replace=False)
y_cols_unseen = [y_col for y_col in y_cols if y_col not in y_cols_seen]
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)
# Callbacks
filepath = 'data/BiLSTM_' + dataset + "_" + str(proportion) + '_' + str(number) + '.h5'
checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=0,
save_best_only=True, mode='auto', save_weights_only=False)
early_stop = EarlyStopping(monitor='val_loss', patience=5, mode='auto')
callbacks_list = [checkpoint, early_stop]
model = BiLSTM(MAX_SEQ_LEN, MAX_FEATURES, EMBEDDING_DIM, n_class_seen, 'model.png', embedding_matrix)
history = model.fit(train_data[0], train_data[1], epochs=30, batch_size=128,
validation_data=valid_data, shuffle=True, verbose=2, callbacks=callbacks_list)
# Save random y_cols for evaluation
d = {'y_cols_seen': list(y_cols_seen),
'y_cols_unseen': list(y_cols_unseen)}
with open('data/y_cols_' + dataset + "_" + str(proportion) + '_' + str(number) + '.pickle', 'wb') as handle:
pickle.dump(d, handle, protocol=pickle.HIGHEST_PROTOCOL)
# Delete cached novelty detection models
try:
os.remove('data/lof_' + dataset + "_" + str(proportion) + '_' + str(number) + '.pickle')
except OSError:
pass
try:
os.remove('data/ocsvm_' + dataset + "_" + str(proportion) + '_' + str(number) + '.pickle')
except OSError:
pass
del le, model, early_stop, checkpoint