forked from mihahauke/dqn_vizdoom_theano
-
Notifications
You must be signed in to change notification settings - Fork 0
/
learn.py
executable file
·167 lines (139 loc) · 4.75 KB
/
learn.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
#!/usr/bin/python
import sys
from time import time
from tqdm import trange
from agents import *
from util import *
setup = dqn_predict
training_steps_per_epoch = 200000
test_episodes_per_epoch = 300
save_params = True
save_results = True
epochs = np.inf
config_loadfile = None
results_loadfile = None
params_loadfile = None
load_params = False
load_results = False
if len(sys.argv) > 3:
load_params = True
load_results = True
params_loadfile = sys.argv[1]
results_loadfile = sys.argv[2]
config_loadfile = sys.argv[3]
params_savefile = params_loadfile
results_savefile = results_loadfile
if load_params:
game = initialize_doom(config_loadfile, True)
engine = QEngine.load(game, params_loadfile)
else:
game, engine = setup()
basefile = engine.name
params_savefile = "params/" + basefile
results_savefile = "results/" + basefile + ".res"
results = None
epoch = 0
if load_results:
results = pickle.load(open(results_loadfile, "r"))
epoch = results["epoch"][-1] + 1
else:
if save_results:
results = dict()
results["epoch"] = []
results["time"] = []
results["overall_time"] = []
results["mean"] = []
results["std"] = []
results["max"] = []
results["min"] = []
results["epsilon"] = []
results["training_episodes_finished"] = []
results["loss"] = []
results["setup"] = engine.setup
print "\nNetwork architecture:"
for p in get_all_param_values(engine.get_network()):
print p.shape
print "\nEngine setup:"
for k in engine.setup.keys():
if k == "network_args":
print"network_args:"
net_args = engine.setup[k]
for k2 in net_args.keys():
print "\t",k2,":",net_args[k2]
else:
print k,":",engine.setup[k]
print
print "============================"
test_frequency = 1
overall_start = time()
if results_loadfile and len(results["time"]) > 0:
overall_start -= results["overall_time"][-1]
# Training starts here!
while epoch < epochs:
print "\nEpoch", epoch
train_time = 0
train_episodes_finished = 0
mean_loss = 0
if training_steps_per_epoch > 0:
rewards = []
start = time()
engine.new_episode(update_state=True)
print "\nTraining ..."
for step in trange(training_steps_per_epoch):
if game.is_episode_finished():
r = game.get_total_reward()
rewards.append(r)
engine.new_episode(update_state=True)
train_episodes_finished += 1
engine.make_learning_step()
end = time()
train_time = end - start
print train_episodes_finished, "training episodes played."
print "Training results:"
print engine.get_actions_stats(clear=True).reshape([-1, 4])
mean_loss = engine._evaluator.get_mean_loss()
if len(rewards) == 0:
rewards.append(-123)
rewards = np.array(rewards)
print "mean:", rewards.mean(), "std:", rewards.std(), "max:", rewards.max(), "min:", rewards.min(), "mean_loss:", mean_loss, "eps:", engine.get_epsilon()
print "t:", sec_to_str(train_time)
# learning mode off
if (epoch + 1) % test_frequency == 0 and test_episodes_per_epoch > 0:
engine.learning_mode = False
rewards = []
start = time()
print "Testing..."
for test_episode in trange(test_episodes_per_epoch):
r = engine.run_episode()
rewards.append(r)
end = time()
print "Test results:"
print engine.get_actions_stats(clear=True, norm=False).reshape([-1, 4])
rewards = np.array(rewards)
print "mean:", rewards.mean(), "std:", rewards.std(), "max:", rewards.max(), "min:", rewards.min()
print "t:", sec_to_str(end - start)
overall_end = time()
overall_time = overall_end - overall_start
if save_results:
print "Saving results to:", results_savefile
results["epoch"].append(epoch)
results["time"].append(train_time)
results["overall_time"].append(overall_time)
results["mean"].append(rewards.mean())
results["std"].append(rewards.std())
results["max"].append(rewards.max())
results["min"].append(rewards.min())
results["epsilon"].append(engine.get_epsilon())
results["training_episodes_finished"].append(train_episodes_finished)
results["loss"].append(mean_loss)
res_f = open(results_savefile, 'w')
pickle.dump(results, res_f)
res_f.close()
epoch += 1
print ""
if save_params:
engine.save(params_savefile)
print "Elapsed time:", sec_to_str(overall_time)
print "========================="
overall_end = time()
print "Elapsed time:", sec_to_str(overall_end - overall_start)