-
Notifications
You must be signed in to change notification settings - Fork 2
/
transductor_train.py
179 lines (142 loc) · 5.9 KB
/
transductor_train.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
'''
run this script until enough transductor models get dumped.
'enough' is any number reasonable for you.
quality increases with quantity, I dumped 300, but 1-2 is okay too
'''
import numpy as np
import pandas as pd
import misc as pt
from misc import add_features
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.utils import np_utils
from sklearn.preprocessing import StandardScaler
from evaluation import roc_auc_truncated, compute_ks, compute_cvm
import cPickle
from scipy.optimize import minimize
import logging
np.random.seed(1337) # for reproducibility
logger = logging.getLogger()
hdlr = logging.FileHandler(pt.transductor_log_file)
formatter = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
hdlr.setFormatter(formatter)
logger.addHandler(hdlr)
logger.setLevel(logging.INFO)
def preprocess_data(X, scaler=None):
if not scaler:
scaler = StandardScaler()
scaler.fit(X[:,:-1]) # don't scale last col - prediction
X[:,:-1] = scaler.transform(X[:,:-1])
return X, scaler
def load(data_file, prediction_file, tail=None, weight=False, mass=False):
data = pd.read_csv(data_file)
data = add_features(data)
prediction = pd.read_csv(prediction_file)
data['prediction'] = prediction["prediction"]
if tail is not None:
data = data[-tail:]
# shuffle
data = data.iloc[np.random.permutation(len(data))].reset_index(drop=True)
filter_out = ['id', 'min_ANNmuon', 'production', 'mass', 'weight', 'signal']
features = list(f for f in data.columns if f not in filter_out)
X = data[features].values
y = data['signal'].values if not mass else None
w = data['weight'].values if weight else None
m = data['mass'].values if mass else None
return X, y, w, m
def create_model(input_shape):
np.random.seed(11) # for reproducibility
model = Sequential()
model.add(Dense(input_shape, 50))
model.add(Activation('tanh'))
model.add(Dense(50, 50))
model.add(Activation('tanh'))
model.add(Dense(50, 30))
model.add(Activation('tanh'))
model.add(Dense(30, 25))
model.add(Activation('tanh'))
model.add(Dense(25, 2))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
return model
def get_weights(model):
weights = model.get_weights()
return np.concatenate([x.ravel() for x in weights])
def set_weights(model, parameters):
weights = model.get_weights()
start = 0
for i,w in enumerate(weights):
size = w.size
weights[i] = parameters[start:start+size].reshape(w.shape)
start += size
model.set_weights(weights)
return model
def dump_transductor_model(model, transductor_model_file):
with open(transductor_model_file, 'wb') as fid:
cPickle.dump(model, fid)
def create_objective(model, transductor_model_file, X, y, Xa, ya, wa, Xc, mc,
ks_threshold=0.09, cvm_threshold=0.002, verbose=True):
i = []
d = []
auc_log = [0]
def objective(parameters):
i.append(0)
set_weights(model, parameters)
p = model.predict(X, batch_size=256, verbose=0)[:, 1]
auc = roc_auc_truncated(y, p)
pa = model.predict(Xa, batch_size=256, verbose=0)[:, 1]
ks = compute_ks(pa[ya == 0], pa[ya == 1], wa[ya == 0], wa[ya == 1])
pc = model.predict(Xc, batch_size=256, verbose=0)[:, 1]
cvm = compute_cvm(pc, mc)
ks_importance = 1 # relative KS importance
ks_target = ks_threshold
cvm_importance = 1 # relative CVM importance
cvm_target = cvm_threshold
alpha = 0.001 # LeakyReLU
ks_loss = (1 if ks > ks_target else alpha) * (ks - ks_target)
cvm_loss = (1 if cvm > cvm_target else alpha) * (cvm - cvm_target)
loss = -auc + ks_importance*ks_loss + cvm_importance*cvm_loss
if ks < ks_threshold and cvm < cvm_threshold and auc > auc_log[0]:
d.append(0)
dump_transductor_model(model, transductor_model_file.format(len(d)))
auc_log.pop()
auc_log.append(auc)
message = "iteration {:7}: Best AUC={:7.5f} achieved, KS={:7.5f}, CVM={:7.5f}".format(len(i), auc, ks, cvm)
logger.info(message)
if verbose:
print("iteration {:7}: AUC: {:7.5f}, KS: {:7.5f}, CVM: {:7.5f}, loss: {:8.5f}".format(len(i),
auc, ks, cvm, loss))
return loss
return objective
Xt, yt, _, _ = load(pt.training_file, pt.train_prediction_file) # shuffled
Xa, ya, wa, _ = load(pt.check_agreement_file, pt.check_agreement_prediction_file,
tail=len(yt), weight=True)
Xc, yc, _, mc = load(pt.check_correlation_file, pt.check_correlation_prediction_file,
mass=True)
Xt, scaler = preprocess_data(Xt)
Xa = preprocess_data(Xa, scaler)[0]
Xc = preprocess_data(Xc, scaler)[0]
with open(pt.transductor_scaler_file, 'wb') as fid:
cPickle.dump(scaler, fid)
AUC = roc_auc_truncated(yt, Xt[:,-1])
print ('AUC before transductor', AUC)
model = create_model(Xt.shape[1])
pretrain = True
if pretrain:
# pretrain model
print("Pretrain model")
yt_categorical = np_utils.to_categorical(yt, nb_classes=2)
model.fit(Xt, yt_categorical, batch_size=64, nb_epoch=1,
validation_data=None, verbose=2, show_accuracy=True)
print("Save pretrained model")
with open(pt.transductor_pretrained_model_file, 'wb') as fid:
cPickle.dump(model, fid)
else:
print("Load pretrained model")
with open(pt.transductor_pretrained_model_file, 'rb') as fid:
model = cPickle.load(fid)
x0 = get_weights(model)
print("Optimize %d weights" % len(x0))
objective = create_objective(model, pt.transductor_model_file,
Xt, yt, Xa, ya, wa, Xc, mc, verbose=True)
minimize(objective, x0, args=(), method='Powell')