-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_tree_model.py
314 lines (248 loc) · 12.2 KB
/
run_tree_model.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
"""
This is the main file to run our approach, DISC, and the full-tree DFA learning method called DFA-FT in our paper.
"""
import read_traces, DFA_utils_tree_minerror, DFA_utils_tree_only, time
import tree_only, tree_minerror
import argparse
from GLOBAL_VARS import *
def run_multigoal_tree_only(base, iteration, timeout=180):
"""
This code learns an automaton for every goal and test its performance.
It reports the convergence performance in the training and testing set.
It also reports the performance of the always-yes and always-no baselines.
Parameters
-------
base: string
The base file prefix of the training and test sets. (e.g. "alfred")
iteration: string
The unique identifier of the iteration for the train/test dataset (e.g. "1", "2", ...)
timeout: int
Time limit (in minutes) given to the model to learn the automaton
"""
path_trainset = "traces/" + base + "/train_" + iteration + ".txt" # Path to the file with the training traces
path_testset = "traces/" + base + "/test_" + iteration + ".txt"
# learning one DFA per goal
if VALIDATION:
G, Sigma, T_train, T_validation = read_traces.read_data_split(path_trainset)
else:
G, Sigma, T_train = read_traces.read_data(path_trainset)
T_validation = []
_, _, T_test = read_traces.read_data(path_testset)
retrain_model = False
avg_num_states = 0
avg_num_transitions = 0
DFAs = {}
t_init = time.time()
goal_priors = DFA_utils_tree_only.goal_priors(T_train + T_validation)
for g_pos in G:
print("Working on goal", g_pos)
q_0, dfa, pos_prob = tree_only.solve_tree_only(g_pos, G, Sigma, T_train, timeout, {})
if VALIDATION:
confidences = DFA_utils_tree_only.test_binary_accuracy(q_0, dfa, pos_prob, T_validation, g_pos)
else:
confidences = DFA_utils_tree_only.test_binary_accuracy(q_0, dfa, pos_prob, T_test, g_pos, smoothing=False)
print("Confidence:", confidences)
print("Goal:", g_pos)
# Saving the automata
DFAs[g_pos] = DFA_utils_tree_only.DFA(q_0, dfa, confidences, pos_prob, g_pos, goal_priors)
sz = DFA_utils_tree_only.dfa_size(q_0, dfa, pos_prob)
avg_num_states += sz[0]
avg_num_transitions += sz[1]
t_end = time.time()
# Checking the train and test performance
# Both values might be lower than 100%
print("----------------------------------------")
if VALIDATION:
conv_train = DFA_utils_tree_only.test_multigoal_convergence_bayes(DFAs, T_train + T_validation)[0]
conv_test = DFA_utils_tree_only.test_multigoal_convergence_bayes(DFAs, T_test)[0]
conv_percent_test = DFA_utils_tree_only.test_multigoal_percent_convergence_bayes(DFAs, T_test)
avg_utility = 0
else:
conv_train = DFA_utils_tree_only.test_multigoal_convergence(DFAs, T_train)
conv_test = DFA_utils_tree_only.test_multigoal_convergence(DFAs, T_test)
conv_percent_test = DFA_utils_tree_only.test_multigoal_percent_convergence(DFAs, T_test)
avg_utility = 0
conv_train = ["%0.3f"%v for v in conv_train]
print(f"Train convergence performance: {conv_train}")
print("----------------------------------------")
conv_test = ["%0.3f"%v for v in conv_test]
print(f"Test convergence performance: {conv_test}")
conv_percent_test = ["%0.3f"%v for v in conv_percent_test]
print(f"Test convergence at 20%: {conv_percent_test[0]}")
print(f"Test convergence at 40%: {conv_percent_test[1]}")
print(f"Test convergence at 60%: {conv_percent_test[2]}")
print(f"Test convergence at 80%: {conv_percent_test[3]}")
print(f"Test convergence at 100%: {conv_percent_test[4]}")
most_common = DFA_utils_tree_only.get_most_common_performance(T_test)
len([1 for tau,g in T_test if g==g_pos])/float(len(T_test))
print(f"Most-Common-Class baseline: {most_common:0.3f}")
print(f"Random classifier baseline: {1/float(len(G)):0.3f}")
print("----------------------------------------")
print(f"Time: {(t_end-t_init)/60:0.2f}[m]")
print("Average num states:", avg_num_states / len(DFAs))
print("Average num edges:", avg_num_transitions / len(DFAs))
outf = open("traces/" + base + "/results/" + "tree_only_" + iteration + ".txt", "w")
outf.write(", ".join(conv_test) + "; " + ", ".join(conv_percent_test) + "; " + str(avg_utility))
outf.close()
sz_f = open("traces/" + base + "/results/" + "tree_only_" + iteration + "_sizes.txt", "w")
sz_f.write("%0.3f, %0.3f\n" % (avg_num_states / len(DFAs), avg_num_transitions / len(DFAs)))
sz_f.close()
def run_multigoal_tree_minerror(base, iteration, q_max, timeout=180):
"""
This code learns an automaton for every goal and test its performance.
It reports the convergence performance in the training and testing set.
It also reports the performance of the always-yes and always-no baselines.
Parameters
-------
base: string
The base file prefix of the training and test sets. (e.g. "alfred")
iteration: string
The unique identifier of the iteration for the train/test dataset (e.g. "1", "2", ...)
q_max: int
The maximum size for any automaton.
timeout: int
Time limit (in minutes) given to the model to learn the automaton
"""
path_trainset = "traces/" + base + "/train_" + iteration + ".txt" # Path to the file with the training traces
path_testset = "traces/" + base + "/test_" + iteration + ".txt"
if VALIDATION:
G, Sigma, T_train, T_validation = read_traces.read_data_split(path_trainset)
else:
G, Sigma, T_train = read_traces.read_data(path_trainset)
T_validation = []
_, _, T_test = read_traces.read_data(path_testset)
DFAs = {}
t_init = time.time()
for g_pos in G:
print("Working on goal", g_pos)
best_hyperparams = None
best_fscore = -1
best_stats = None
best_model = None
for loop_penalty in [0.001]:
for trans_penalty in [10, 3, 1, 0.3, 0.1, 0.03, 0.01, 0.003, 0.001, 0.0003, 0.0001]:
q_0, dfa, is_q0_pos, always_returns_no, goal_priors = tree_minerror.solve_tree_minerror(g_pos, q_max, G, Sigma, T_train, timeout, {}, loop_penalty = loop_penalty, trans_penalty = trans_penalty)
if VALIDATION:
valid_stats = DFA_utils_tree_minerror.test_binary_accuracy(q_0, dfa, T_validation, g_pos, is_q0_pos, always_returns_no)
else:
valid_stats = DFA_utils_tree_minerror.test_binary_accuracy(q_0, dfa, T_train, g_pos, is_q0_pos, always_returns_no)
if valid_stats[5] > best_fscore:
best_fscore = valid_stats[5]
best_hyperparams = (loop_penalty, trans_penalty)
#best_conv = conv
best_stats = valid_stats
best_model = (q_0, dfa, is_q0_pos, always_returns_no, goal_priors)
print(best_hyperparams)
q_0, dfa, is_q0_pos, always_returns_no, goal_priors = best_model
confidence = best_stats
# Check if the DFA always returns no for any possible trace -- if so, modify the confidences accordingly.
if always_returns_no:
confidence = (confidence[0], confidence[1], 0, 1, confidence[4], 0)
test_acc = DFA_utils_tree_minerror.test_binary_convergence(q_0, dfa, T_test, g_pos, is_q0_pos)
print("Goal:", g_pos, "Prior prob: ", goal_priors[g_pos])
print("Accuracy/Precision/Recall/TNR/FNR/F2-score:", confidence[0], confidence[1],confidence[2], confidence[3], confidence[4], confidence[5])
print("Real test accuracy:", test_acc)
test_conf = DFA_utils_tree_minerror.test_binary_accuracy(q_0, dfa, T_test, g_pos, is_q0_pos, always_returns_no, smoothing = False)
if always_returns_no:
test_conf = (test_conf[0], test_conf[1], 0, 1)
print("Real test accuracy/precision/recall/TNR", test_conf[0], test_conf[1], test_conf[2], test_conf[3])
# Saving the automata
DFA_utils_tree_minerror.show_dfa(q_0, dfa, is_q0_pos)
DFAs[g_pos] = DFA_utils_tree_minerror.DFA(q_0, dfa, confidence, g_pos, is_q0_pos, goal_priors)
t_end = time.time()
# Checking the train and test performance
# Both values might be lower than 100%
print("----------------------------------------")
if VALIDATION:
conv_train, conv_train_prob = DFA_utils_tree_minerror.test_multigoal_convergence(DFAs, T_train + T_validation)
else:
conv_train, conv_train_prob = DFA_utils_tree_minerror.test_multigoal_convergence(DFAs, T_train)
conv_train = ["%0.3f"%v for v in conv_train]
conv_train_prob = ["%0.3f"%v for v in conv_train_prob]
print(f"Train convergence argmax performance: {conv_train}")
print(f"Train convergence probability performance: {conv_train_prob}")
print("----------------------------------------")
conv_test, conv_test_prob = DFA_utils_tree_minerror.test_multigoal_convergence(DFAs, T_test, topk=1)
conv_test = ["%0.3f"%v for v in conv_test]
conv_test_prob = ["%0.3f"%v for v in conv_test_prob]
print(f"Test convergence argmax performance: {conv_test}")
print(f"Test convergence probability performance: {conv_test_prob}")
print("----------------------------------------")
conv_percent_test = DFA_utils_tree_minerror.test_multigoal_percent_convergence(DFAs, T_test, topk=1)
conv_percent_test = ["%0.3f"%v for v in conv_percent_test]
print(f"Test convergence at 20%: {conv_percent_test[0]}")
print(f"Test convergence at 40%: {conv_percent_test[1]}")
print(f"Test convergence at 60%: {conv_percent_test[2]}")
print(f"Test convergence at 80%: {conv_percent_test[3]}")
print(f"Test convergence at 100%: {conv_percent_test[4]}")
print("----------------------------------------")
avg_utility = DFA_utils_tree_minerror.test_early_utility(DFAs, T_test)
print("----------------------------------------")
most_common = DFA_utils_tree_minerror.get_most_common_performance(T_test)
len([1 for tau,g in T_test if g==g_pos])/float(len(T_test))
print(f"Most-Common-Class baseline: {most_common:0.3f}")
print(f"Random classifier baseline: {1/float(len(G)):0.3f}")
print("----------------------------------------")
print(f"Time: {(t_end-t_init)/60:0.2f}[m]")
outf = open("traces/" + base + "/results/" + "ours_" + iteration + ".txt", "w")
outf.write(", ".join(conv_test) + "; " + ", ".join(conv_percent_test) + "; " + str(avg_utility))
outf.close()
modelf = open("traces/" + base + "/results/" + "ours_" + iteration + "_models.txt", "w")
for g in DFAs:
modelf.write(DFA_utils_tree_minerror.report_dfa(DFAs[g].q_0, DFAs[g].dfa, g))
modelf.close()
def run_multilabel_tree_minerror(base, iteration, q_max, timeout=180):
"""
This code is the version of run_tree_minerror for the multilabel experiment.
"""
assert(MULTILABEL)
path_trainset = "traces/" + base + "/train_" + iteration + ".txt" # Path to the file with the training traces
path_testset = "traces/" + base + "/test_" + iteration + ".txt"
G, Sigma, T_train = read_traces.read_data(path_trainset)
_, _, T_test = read_traces.read_data(path_testset)
DFAs = {}
for g_pos in G:
q_0, dfa, is_q0_pos, always_returns_no, goal_priors = tree_minerror.solve_tree_minerror(g_pos, q_max, G, Sigma, T_train, timeout, {}, loop_penalty = 0.001, trans_penalty = 0.0003)
# Saving the automata
DFA_utils_tree_minerror.show_dfa(q_0, dfa, is_q0_pos)
DFAs[g_pos] = DFA_utils_tree_minerror.DFA(q_0, dfa, None, g_pos, is_q0_pos, goal_priors)
# Checking the train and test performance
# Both values might be lower than 100%
print("----------------------------------------")
conv_train = DFA_utils_tree_minerror.test_multilabel_convergence(DFAs, T_train)
conv_train = ["%0.3f"%v for v in conv_train]
print(f"Train convergence performance: {conv_train}")
print("----------------------------------------")
conv_test = DFA_utils_tree_minerror.test_multilabel_convergence(DFAs, T_test)
conv_test = ["%0.3f"%v for v in conv_test]
print(f"test convergence performance: {conv_test}")
outf = open("traces/" + base + "/results/" + "ours_" + iteration + ".txt", "w")
outf.write(", ".join(conv_test))
outf.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', help='e.g. alfred, crystal')
parser.add_argument('--qmax', type=int)
parser.add_argument('--id', type=int)
args = parser.parse_args()
base = args.dataset
q_max = args.qmax
iteration = args.id
# Default Parameters
if base is None:
base = "alfred"
if q_max is None:
q_max = 10
if iteration is None:
iteration = 1
iteration = str(iteration)
print("traces/" + base + "/train_" + iteration + ".txt")
if TREE_MINERROR:
if MULTILABEL:
run_multilabel_tree_minerror(base, iteration, q_max, timeout=15)
else:
run_multigoal_tree_minerror(base, iteration, q_max, timeout=15)
else:
run_multigoal_tree_only(base, iteration, timeout=15)
print("traces/" + base + "/train_" + iteration + ".txt")
## mit: (0.01, 0.5)