forked from koulanurag/mmn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
fsm_process.py
167 lines (145 loc) · 8.72 KB
/
fsm_process.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
"""
Generating data, training and testing functions are implemented here.
"""
import os
import qbn
import time
import copy
import torch
import pickle
import gru_nn
import bgru_nn
import logging
import tools as tl
from torch import optim
from moore_machine import MooreMachine
import ipdb
import sys
class ProcessFSM():
def __init__(self, env):
self.env = env
def generate_train_data(self, no_batches, batch_size, trajectories_data_path, generate_train_data, gru_dir):
tl.set_log(gru_dir, 'generate_train_data')
train_data = tl.generate_trajectories(self.env, no_batches, batch_size, trajectories_data_path)
return train_data
def train_gru(self, gru_net, gru_net_path, gru_plot_dir, train_data, batch_size, train_epochs, cuda, bn_episodes, bottleneck_data_path, generate_max_steps, gru_prob_data_path, gru_dir):
logging.info('Training GRU!')
start_time = time.time()
gru_net.train()
optimizer = optim.Adam(gru_net.parameters(), lr=1e-3)
gru_net = gru_nn.train(gru_net, self.env, optimizer, gru_net_path, gru_plot_dir, train_data, batch_size,
train_epochs, cuda, trunc_k=50)
logging.info('Generating Data-Set for Later Bottle Neck Training')
gru_net.eval()
tl.generate_bottleneck_data(gru_net, self.env, bn_episodes, bottleneck_data_path, cuda=cuda, max_steps=generate_max_steps)
tl.generate_trajectories(self.env, 500, batch_size, gru_prob_data_path, gru_net.cpu())
tl.write_net_readme(gru_net, gru_dir, info={'time_taken': time.time() - start_time})
return gru_net
def test_gru(self, trained_gru, gru_net_path, cuda):
logging.info('Testing GRU!')
trained_gru.load_state_dict(torch.load(gru_net_path))
trained_gru.eval()
trained_gru.noise = False
no_episodes = 20
perf = gru_nn.test(trained_gru, self.env, no_episodes, log=True, cuda=cuda, render=True)
logging.info('Average Performance:{}'.format(perf))
return perf
def bhx_train(self, bhx_net, hx_train_data, hx_test_data, bhx_net_path, bhx_plot_dir, batch_size, train_epochs, cuda, target_net, bhx_dir):
bhx_start_time = time.time()
logging.info('Training HX SandGlassNet!')
optimizer = optim.Adam(bhx_net.parameters(), lr=1e-4, weight_decay=0)
bhx_net.train()
bhx_net = qbn.train(bhx_net, (hx_train_data, hx_test_data), optimizer, bhx_net_path, bhx_plot_dir,
batch_size, train_epochs, cuda, grad_clip=5, target_net=target_net, env=self.env,
low=-0.02, high=0.02)
bhx_end_time = time.time()
tl.write_net_readme(bhx_net, bhx_dir, info={'time_taken': round(bhx_end_time - bhx_start_time, 4)})
def bhx_test(self, bhx_net, bhx_net_path, hx_test_data, cuda):
logging.info('Testing HX SandGlassNet')
bhx_net.load_state_dict(torch.load(bhx_net_path))
bhx_net.eval()
bhx_test_mse = qbn.test(bhx_net, hx_test_data, len(hx_test_data), cuda=cuda)
logging.info('MSE :{}'.format(bhx_test_mse))
def ox_train(self, ox_net, obs_train_data, obs_test_data, ox_net_path, ox_plot_dir, batch_size, train_epochs, cuda, target_net, ox_dir):
ox_start_time = time.time()
logging.info('Training OX SandGlassNet!')
optimizer = optim.Adam(ox_net.parameters(), lr=1e-4, weight_decay=0)
ox_net.train()
ox_net = qbn.train(ox_net, (obs_train_data, obs_test_data), optimizer, ox_net_path, ox_plot_dir,
batch_size, train_epochs, cuda, grad_clip=5, target_net=target_net, env=self.env,
low=-0.02, high=0.02)
ox_end_time = time.time()
tl.write_net_readme(ox_net, ox_dir, info={'time_taken': round(ox_end_time - ox_start_time, 4)})
def ox_test(self, ox_net, ox_net_path, obs_test_data, cuda):
logging.info('Testing OX SandGlassNet')
ox_net.load_state_dict(torch.load(ox_net_path))
ox_net.eval()
ox_test_mse = qbn.test(ox_net, obs_test_data, len(obs_test_data), cuda=cuda)
logging.info('MSE : {}'.format(ox_test_mse))
def bgru_train(self, bgru_net, gru_net, cuda, gru_scratch, trajectories_data_path, bgru_net_path, bgru_plot_dir, batch_size, train_epochs, gru_prob_data_path, bgru_dir):
self.env.spec.reward_threshold = gru_nn.test(gru_net, self.env, 10, log=True, cuda=cuda, render=True)
logging.info('Training Binary GRUNet!')
bgru_net.train()
_start_time = time.time()
if gru_scratch:
optimizer = optim.Adam(bgru_net.parameters(), lr=1e-3)
train_data = tl.generate_trajectories(self.env, 3, 5, trajectories_data_path)
bgru_net = gru_nn.train(bgru_net, self.env, optimizer, bgru_net_path, bgru_plot_dir, train_data, batch_size,
train_epochs, cuda)
else:
optimizer = optim.Adam(bgru_net.parameters(), lr=1e-4)
train_data = tl.generate_trajectories(self.env, 3, 5, gru_prob_data_path, copy.deepcopy(bgru_net.gru_net).cpu())
bgru_net = bgru_nn.train(bgru_net, self.env, optimizer, bgru_net_path, bgru_plot_dir, train_data, 5,
train_epochs, cuda, test_episodes=1, trunc_k=100)
tl.write_net_readme(bgru_net, bgru_dir, info={'time_taken': round(time.time() - _start_time, 4)})
def bgru_test(self, bgru_net, bgru_net_path, cuda):
bgru_net.load_state_dict(torch.load(bgru_net_path))
bgru_net.eval()
bgru_perf = bgru_nn.test(bgru_net, self.env, 1, log=True, cuda=cuda, render=True)
logging.info('Average Performance: {}'.format(bgru_perf))
def generate_fsm(self, bgru_net, bgru_net_path, cuda, bgru_dir, full_table):
bgru_net.load_state_dict(torch.load(bgru_net_path))
bgru_net.eval()
moore_machine = MooreMachine()
# TODO: Full transaction table은 구할 수 없다. notion 참조.
# if full_table:
# moore_machine.extract_from_nn(self.env, bgru_net, 2, 0, log=True, partial=False, cuda=cuda)
# pickle.dump(moore_machine, open(os.path.join(bgru_dir, 'full_unmin_moore_machine.p'), 'wb'))
# moore_machine.save(open(os.path.join(bgru_dir, 'full_fsm.txt'), 'w'))
# # minimize_full_fsm
# moore_machine.minimize()
# moore_machine.save(open(os.path.join(bgru_dir, 'full_minimized_moore_machine.txt'), 'w'))
# pickle.dump(moore_machine, open(os.path.join(bgru_dir, 'full_min_moore_machine.p'), 'wb'))
# else: # original mmn
moore_machine.extract_from_nn(self.env, bgru_net, 1, 0, log=True, partial=True, cuda=cuda)
pickle.dump(moore_machine, open(os.path.join(bgru_dir, 'partial_unmin_moore_machine.p'), 'wb'))
moore_machine.save(open(os.path.join(bgru_dir, 'partial_fsm.txt'), 'w'))
# minimize_partial_fsm
moore_machine.minimize_partial_fsm(bgru_net)
moore_machine.save(open(os.path.join(bgru_dir, 'partial_minimized_moore_machine.txt'), 'w'))
pickle.dump(moore_machine, open(os.path.join(bgru_dir, 'partial_min_moore_machine.p'), 'wb'))
def evaluate_fsm(self, bgru_net, bgru_net_path, bgru_dir, full_table):
bgru_net.load_state_dict(torch.load(bgru_net_path))
# if not full_table:
# # TODO: full transaction table을 구성할 수 없음을 알았다.
# print('we cannot evaluate partial_minimzed_fsm !')
# sys.exit()
# moore_machine = pickle.load(open(os.path.join(bgru_dir, 'full_min_moore_machine.p)'), 'rb'))
moore_machine = pickle.load(open(os.path.join(bgru_dir, 'partial_min_moore_machine.p)'), 'rb'))
bgru_net.cpu()
bgru_net.eval()
perf = moore_machine.evaluate(bgru_net, self.env, total_episodes=3, render=True, inspect=False)
logging.info('Moore Machine Performance: {}'.format(perf))
def functional_pruning_fsm(self, bgru_net, bgru_dir, cuda):
bgru_net_path = os.path.join(bgru_dir, 'model.p')
bgru_plot_dir = tl.ensure_directory_exits(os.path.join(bgru_dir, 'Plots'))
# TODO: unminimized, partial MMN을 pruning 해야한다.
# min_moore_machine_path = os.path.join(bgru_dir, 'full_min_moore_machine.p')
# unmin_moore_machine_path = os.path.join(bgru_dir, 'full_unmin_moore_machine.p')
unmin_moore_machine_path = os.path.join(bgru_dir, 'partial_unmin_moore_machine.p')
bgru_net.load_state_dict(torch.load(bgru_net_path))
moore_machine = pickle.load(open(min_moore_machine_path, 'rb'))
bgru_net.eval()
# TODO: Store_obs는 일단 False로 한다
path = bgru_net_path
abcd = moore_machine.functional_pruning(bgru_net, self.env, log=True, store_obs = False, path = bgru_dir)