/
dagger.py
224 lines (175 loc) · 6.53 KB
/
dagger.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
import random
import sys
import numpy as np
import scipy.sparse as spa
from sklearn.linear_model import SGDClassifier
from sklearn.svm import LinearSVC, SVC
from sklearn.neighbors import DistanceMetric
from sklearn.metrics import classification_report
from sklearn.model_selection import KFold
from utils import readDataset, Processor, Sequencer, test, save
class Dagger(object):
def __init__(self, proc, Xss, yss, valid_set=0.1, validation_set=None):
self.seen_states = set()
self.state_set = []
self.proc = proc
self.valid_set = 0.1
self.surr_loss = DistanceMetric.get_metric('hamming')
if validation_set is None:
self._split(Xss, yss)
else:
self.Xss = Xss
self.yss = yss
self.valid_Xss, self.valid_yss = validation_set
def _split(self, Xss, yss):
train_Xss, valid_Xss = [], []
train_yss, valid_yss = [], []
for Xs, ys in zip(Xss, yss):
if np.random.binomial(1, 1 - self.valid_set) == 1:
train_Xss.append(Xs)
train_yss.append(ys)
else:
valid_Xss.append(Xs)
valid_yss.append(ys)
self.Xss = train_Xss
self.yss = train_yss
self.valid_Xss = valid_Xss
self.valid_yss = valid_yss
def generate(self, Xs, ys, exp, sequencer):
"""
Generates a new tradjectory.
"""
# copy orig seq
trad = []
for i in range(len(Xs)):
if np.random.random() < exp:
# Oracle
output = ys[i]
else:
# Policy
output = sequencer._partial_pred(Xs, trad, i)
trad.append(output)
return trad
def gen_dataset(self, clf, epoch):
sequencer = Sequencer(self.proc, clf)
# Generate new dataset
for Xs, ys in zip(self.Xss, self.yss):
# build tradjectory
trad = self.generate(Xs, ys, 1 if epoch == 0 else 0, sequencer)
# If different than the input, add it
if any(y != t for y, t in zip(ys, trad)):
self.add_sequence(Xs, ys, trad)
#pprint.pprint(zip([X['feature'] for X in Xs], ys, trad))
def add_sequence(self, Xs, ys, trads, force=False):
for i in range(len(Xs)):
state = self.get_state(Xs, trads, i)
if state not in self.seen_states or force:
X = self.proc.transform(Xs, trads, i)
y = self.proc.encode_target(ys, i)[0]
self.state_set.append((X, y))
self.seen_states.add(state)
def get_state(self, Xs, trad, idx):
return ' '.join(self.proc.state(Xs, trad, idx))
def score_policy(self, clf):
sequencer = Sequencer(self.proc, clf)
scores = []
for Xs, ys in zip(self.valid_Xss, self.valid_yss):
trads = [i for i in sequencer.classify(Xs, raw=True)]
expected = [self.proc.encode_target(ys, i)[0] for i in range(len(ys))]
scores.append(self.surr_loss.pairwise([trads], [expected]))
return np.mean(scores)
def train(self, epochs=30):
# Add to the initial set
states = 0
for Xs, ys in zip(self.Xss, self.yss):
states += len(ys)
self.add_sequence(Xs, ys, ys, force=True)
print(len(self.seen_states), "unique,", states, "total")
# Initial policy just mimics the expert
clf = self.train_model()
# Get best policy found so far
bscore, bclf= self.score_policy(clf), clf
print("Best score seen:", bscore)
for e in range(1, epochs):
# Generate new dataset
print("Generating new dataset")
dataSize = len(self.state_set)
self.gen_dataset(clf, e)
if dataSize == len(self.state_set):
break
# Retrain
print("Training")
clf = self.train_model()
print("Scoring")
score = self.score_policy(clf)
print("New Policy Score:", bscore)
if score < bscore:
bscore = score
bclf = clf
return bclf
def train_model(self):
print("Featurizing...")
tX, tY = [], []
for X, y in self.state_set:
tX.append(X)
tY.append(y)
tX, tY = spa.vstack(tX), np.vstack(tY)
print("Running learner...")
clf = SGDClassifier(loss="hinge", penalty="l2", max_iter=50, tol=1e-3)
#clf = LinearSVC(penalty="l2", class_weight='auto')
print("Samples:", tX.shape[0])
clf.fit(tX, tY.ravel())
return clf
def subset(Xss, yss, idxs, rs, shuffle=True):
# could be range object
if type(idxs) != list:
idxs = list(idxs)
if shuffle:
rs.shuffle(idxs)
print( "Train IDXS", idxs[:10], "...")
tXss = [Xss[i] for i in idxs]
tyss = [yss[i] for i in idxs]
return tXss, tyss
def main(fn, output_fn):
print("Reading in dataset")
data, classes = readDataset(fn)
print(len(data), " sequences found")
print("Found classes:", sorted(classes))
proc = Processor(classes, 2, 2, prefix=(1,3), affix=(2,1), hashes=2,
features=100000, stem=False, ohe=False)
yss = []
ryss = []
for Xs in data:
ys = [x['output'] for x in Xs]
yss.append(ys)
ryss.append([proc.encode_target(ys, i) for i in range(len(ys))])
rs = np.random.RandomState(seed=2016)
print("Starting KFolding")
y_trues, y_preds = [], []
fold_object = KFold(5, random_state=1)
for train_idx, test_idx in fold_object.split(data):
tr_X, tr_y = subset(data, yss, train_idx, rs)
test_data = subset(data, yss, test_idx, rs, False)
print("Training")
d = Dagger(proc, tr_X, tr_y, validation_set=test_data)
clf = d.train(10)
seq = Sequencer(proc, clf)
print("Testing")
y_true, y_pred = test(data, ryss, test_idx, seq)
# print(y_true, y_pred, proc.labels)
print( classification_report(y_true, y_pred))
y_trues.extend(y_true)
y_preds.extend(y_pred)
print("Total Report")
print(classification_report(y_trues, y_preds, target_names=proc.labels))
print("Training all")
idxs = range(len(data))
tr_X, tr_y = subset(data, yss, idxs, rs)
d = Dagger(proc, tr_X, tr_y)
clf = d.train()
seq = Sequencer(proc, clf)
save(output_fn, seq)
if __name__ == '__main__':
random.seed(0)
np.random.seed(0)
main(sys.argv[1], sys.argv[2])