-
Notifications
You must be signed in to change notification settings - Fork 19
/
Copy pathmain.py
235 lines (206 loc) · 9.91 KB
/
main.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
from env import Ur5
from env2 import Ur5_vision
from DDPG import DDPG
from TD3 import TD3
from TD3_vision import TD3_vision
import numpy as np
import matplotlib.pyplot as plt
import torch
import time
import argparse
import os
def get_time(start_time):
m, s = divmod(int(time.time()-start_time), 60)
h, m = divmod(m, 60)
print 'Total time spent: %d:%02d:%02d' % (h, m, s)
def img_transform(img_memory, mode, frame_size=4):
assert type(img_memory).__module__ == np.__name__, 'data type is not numpy'
assert mode == 'img2txt' or mode == 'txt2img', 'Please use correct mode name 1.img2txt 2.txt2img'
if mode == 'img2txt':
size = img_memory.shape[0]
img_memory = img_memory.reshape((size,-1))
if mode == 'txt2img':
size = img_memory.shape[0]
h = w = np.sqrt(img_memory.shape[1] / (2 * frame_size))
img_memory = img_memory.reshape((size,2,frame_size,h,w))
return img_memory
def train(args, env, model):
if not os.path.exists(args.path_to_model+args.model_name+args.model_date):
os.makedirs(args.path_to_model+args.model_name+args.model_date)
#training reward list
total_reward_list = np.array([])
#testing reward and steps list
test_reward_list, test_step_list = np.array([]), np.array([])
start_time = time.time()
if args.pre_train:
#load pre_trained model
try:
model.load_model(args.path_to_model+args.model_name, args.model_date_+'/')
print 'load model successfully'
except:
print 'fail to load model, check the path of models'
print 'start random exploration for adding experience'
if args.model_name == 'TD3_vision':
vision, state = env.reset()
else:
state = env.reset()
for step in range(args.random_exploration):
if args.model_name == 'TD3_vision':
vision_, state_, action, reward, terminal = env.uniform_exploration(np.random.uniform(-1,1,5)*args.action_bound*5)
model.store_transition(vision,state,action,reward,vision_,state_,terminal)
state = state_
vision = vision_
if terminal:
vision, state = env.reset()
else:
state_, action, reward, terminal = env.uniform_exploration(np.random.uniform(-1,1,5)*args.action_bound*5)
model.store_transition(state,action,reward,state_,terminal)
state = state_
if terminal:
state = env.reset()
total_reward_list = np.loadtxt(args.path_to_model+args.model_name+args.model_date_+'/reward.txt')
test_reward_list = np.loadtxt(args.path_to_model+args.model_name+args.model_date_+'/test_reward.txt')
test_step_list = np.loadtxt(args.path_to_model+args.model_name+args.model_date_+'/test_step.txt')
print 'start training'
model.mode(mode='train')
#training for vision observation
for epoch in range(args.train_epoch):
if args.model_name == 'TD3_vision':
vision, state = env.reset()
else:
state = env.reset()
total_reward = 0
for i in range(args.train_step):
if args.model_name == 'TD3_vision':
action = model.choose_action(vision,state)
vision_, state_, reward, terminal = env.step(action*args.action_bound)
model.store_transition(vision,state,action,reward,vision_,state_,terminal)
state = state_
vision = vision_
total_reward += reward
if model.memory_counter > args.random_exploration:
model.Learn()
if terminal:
vision, state = env.reset()
else:
action = model.choose_action(state)
state_, reward, terminal = env.step(action*args.action_bound)
model.store_transition(state,action,reward,state_,terminal)
state = state_
total_reward += reward
if model.memory_counter > args.random_exploration:
model.Learn()
if terminal:
state = env.reset()
np.append(total_reward_list,total_reward)
print 'epoch:', epoch, '||', 'Reward:', total_reward
#begin testing and record the evalation metrics
if (epoch+1) % args.test_epoch == 0:
model.save_model(args.path_to_model+args.model_name, args.model_date+'/')
model.plot_loss(args.path_to_model+args.model_name, args.model_date+'/')
avg_reward, avg_step = test(args, env, model)
model.mode(mode='train')
print 'finish testing'
np.append(test_reward_list,avg_reward)
np.append(test_step_list,avg_step)
plt.figure()
plt.plot(np.arange(len(test_reward_list)), test_reward_list)
plt.ylabel('test_reward')
plt.xlabel('training epoch / testing epoch')
plt.savefig(args.path_to_model+args.model_name+args.model_date+'/test_reward.png')
plt.close()
np.savetxt(args.path_to_model+args.model_name+args.model_date+'/test_reward.txt',np.array(test_reward_list))
plt.figure()
plt.plot(np.arange(len(test_step_list)), test_step_list)
plt.ylabel('test_step')
plt.xlabel('training epoch / testing epoch')
plt.savefig(args.path_to_model+args.model_name+args.model_date+'/test_step.png')
plt.close()
np.savetxt(args.path_to_model+args.model_name+args.model_date+'/test_step.txt',np.array(test_step_list))
plt.figure()
plt.plot(np.arange(len(total_reward_list)), total_reward_list)
plt.ylabel('Total_reward')
plt.xlabel('training epoch')
plt.savefig(args.path_to_model+args.model_name+args.model_date+'/reward.png')
plt.close()
np.savetxt(args.path_to_model+args.model_name+args.model_date+'/reward.txt',np.array(total_reward_list))
get_time(start_time)
def test(args, env, model):
model.mode(mode='test')
print 'start to test the model'
try:
model.load_model(args.path_to_model+args.model_name, args.model_date_+'/')
print 'load model successfully'
except:
print 'fail to load model, check the path of models'
total_reward_list = []
steps_list = []
#testing for vision observation
for epoch in range(args.test_epoch):
if args.model_name == 'TD3_vision':
vision, state = env.reset()
else:
state = env.reset()
total_reward = 0
for step in range(args.test_step):
if args.model_name == 'TD3_vision':
action = model.choose_action(vision,state,noise=None)
vision_, state_, reward, terminal = env.step(action*args.action_bound)
state = state_
vision = vision_
total_reward += reward
if env.get_rotation > env.threshold:
steps_list.append(env.steps)
if terminal:
vision, state = env.reset()
else:
action = model.choose_action(state,noise=None)
state_, reward, terminal = env.step(action*args.action_bound)
state = state_
total_reward += reward
if terminal:
env.reset()
total_reward_list.append(total_reward)
print 'testing_epoch:', epoch, '||', 'Reward:', total_reward
average_reward = np.mean(np.array(total_reward_list))
average_step = 0 if steps_list == [] else np.mean(np.array(steps_list))
return average_reward, average_step
if __name__ == '__main__':
parser = argparse.ArgumentParser()
#select env to be used
parser.add_argument('--env_name', default='vision')
#select model to be used
parser.add_argument('--model_name', default='TD3_vision')
#Folder name saved as date
parser.add_argument('--model_date', default='/22_01_2019')
#Folder stored with trained model weights, which are used for transfer learning
parser.add_argument('--model_date_', default='/22_01_2019')
parser.add_argument('--pre_train', default=False)
parser.add_argument('--path_to_model', default='/home/waiyang/pana_RL_yueci/')
#The maximum action limit
parser.add_argument('--action_bound', default=np.pi/72, type=float) #pi/36 for reaching
parser.add_argument('--train_epoch', default=500, type=int)
parser.add_argument('--train_step', default=200, type=int)
parser.add_argument('--test_epoch', default=10, type=int)
parser.add_argument('--test_step', default=200, type=int)
#exploration (randome action generation) steps before updating the model
parser.add_argument('--random_exploration', default=1000, type=int)
#store the model weights and plots after epoch number
parser.add_argument('--epoch_store', default=10, type=int)
#Wether to use GPU
parser.add_argument('--cuda', default=False)
parser.add_argument('--mode', default='train')
args = parser.parse_args()
assert args.env_name == 'empty' or 'vision', 'env name: 1.empty 2.vision'
if args.env_name == 'empty': env = Ur5()
if args.env_name == 'vision': env = Ur5_vision()
assert args.model_name == 'TD3_vision' or 'TD3' or 'DDPG', 'model name: 1.TD3_vision 2.TD3 3.DDPG'
if args.model_name == 'TD3_vision': model = TD3_vision(a_dim=env.action_dim,s_dim=env.state_dim,cuda=args.cuda)
if args.model_name == 'TD3': model = TD3(a_dim=env.action_dim,s_dim=env.state_dim,cuda=args.cuda)
if args.model_name == 'DDPG': model = DDPG(a_dim=env.action_dim,s_dim=env.state_dim,cuda=args.cuda)
assert args.mode == 'train' or 'test', 'mode: 1.train 2.test'
if args.mode == 'train':
train(args, env, model)
if args.mode == 'test':
env.duration = 0.1
test(args, env, model)