forked from createamind/candy
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcarla_wrapper.py
338 lines (267 loc) · 10.5 KB
/
carla_wrapper.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
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
# -*- coding: utf-8 -*-
from candy_model import Machine
import numpy as np
import yaml
from carla import image_converter
from carla import sensor
from carla.client import make_carla_client, VehicleControl
from tqdm import tqdm
import msgpack
import msgpack_numpy as m
m.patch()
import os
# BUFFER_LIMIT = 258
BATCH_SIZE = 128
KEEP_CNT = 1500
MAX_SAVE = 0
TRAIN_EPOCH = 30
class Carla_Wrapper(object):
def __init__(self, gamma=0.99, lam=0.95):
self.machine = Machine()
self.global_step = 0
self.lam = lam
self.gamma = gamma
self.state = self.machine.ppo.initial_state
self.obs, self.actions, self.values, self.neglogpacs, self.rewards, self.vaerecons, self.states, self.std_actions, self.manual = [],[],[],[],[],[],[],[],[]
#obs:观察到的东西((连续的两帧包括深度的图像, speed), actions:采取的动作的code, values:PPO中critic的对当前局面的估价, neglogpacs：PPO中actor对当前动作的概率的负log, rewards：当前的奖励
#vaerecons:VAE的重建误差，用于difficulty的计算,进行优先采样 states：ＰＰＯ中ＬＳＴＭ的中间状态, std_actions:人给的动作的ｃｏｄｅ, manual：是否应该使用人给的动作进行模仿学习
self.last_frame = None
# self.pretrain()
def analyze_control(self, control):
steer = control.steer
throttle = control.throttle
brake = control.brake
hand_brake = control.hand_brake
reverse = control.reverse
# b = 0
# if steer < -0.2:
# b += 2
# elif steer > 0.2:
# b += 1
# b *= 3
# if throttle > 0.4:
# b += 2
# if brake > 0.4:
# b += 1
# if reverse:
# b = 9
# if hand_brake:
# if steer < -0.4:
# b = 10
# elif steer > 0.4:
# b = 11
# else:
# b = 12
th = 0
if brake > 0:
th = -brake
if throttle > 0:
th = throttle
return [th, steer]
def process_sensor_data(self, sensor_data):
_main_image = sensor_data.get('CameraRGB', None)
_mini_view_image1 = sensor_data.get('CameraDepth', None)
_mini_view_image2 = sensor_data.get('CameraSemSeg', None)
t1 = np.array(image_converter.to_rgb_array(_main_image)).astype(np.float32) / 128 - 1
t2 = np.max(image_converter.depth_to_logarithmic_grayscale(_mini_view_image1), axis=2, keepdims=True) / 128 - 1
t3 = np.max(image_converter.to_rgb_array(_mini_view_image2), axis=2)
return [t1, t2, t3]
# def update_reward(self, cnt, obs, action):
# _, last_values, _, _, _ = self.machine.value(obs, self.state, action)
# #discount/bootstrap off value fn
# self.advs = np.zeros_like(self.rewards)
# # self.rewards[-1] = reward
# # print(' '.join([('%.2f' % i)for i in self.rewards]))
# l = len(self.obs)
# lastgaelam = 0
# for t in reversed(range(l - cnt, l)):
# if t == l - 1:
# nextvalues = last_values
# else:
# nextvalues = self.values[t+1]
# delta = self.rewards[t] + self.gamma * nextvalues - self.values[t]
# self.advs[t] = lastgaelam = delta + self.gamma * self.lam * lastgaelam
# for t in range(l - cnt, l):
# self.rewards[t] = float(self.advs[t] + self.values[t])
def update_reward(self, cnt, obs, action, reward):
# #discount/bootstrap off value fn
# self.advs = np.zeros_like(self.rewards)
# # self.rewards[-1] = reward
# # print(' '.join([('%.2f' % i)for i in self.rewards]))
# lastgaelam = 0
# for t in reversed(range(l - cnt, l)):
# if t == l - 1:
# nextvalues = last_values
# else:
# nextvalues = self.values[t+1]
# delta = self.rewards[t] + self.gamma * nextvalues - self.values[t]
# self.advs[t] = lastgaelam = delta + self.gamma * self.lam * lastgaelam
l = len(self.obs)
for t in range(l - cnt, l - 2):
self.rewards[t] = self.rewards[t+1]
if reward is None:
self.rewards[l-1] = self.rewards[l-2]
else:
self.rewards[l-1] = reward
self.rewards[l-1] *= 20
for t in reversed(range(l - cnt, l - 1)):
self.rewards[t] += self.lam * self.rewards[t+1]
def post_process(self, inputs, cnt):
obs, reward, action, std_action, manual = self.pre_process(inputs)
#batch of steps to batch of rollouts
# self.obs = np.asarray(self.obs, dtype=np.float32)
# self.actions = np.asarray(self.actions, dtype=np.int32)
# self.values = np.asarray(self.values, dtype=np.float32)
# self.neglogpacs = np.asarray(self.neglogpacs, dtype=np.float32)
# self.rewards = np.asarray(self.rewards, dtype=np.float32)
# self.vaerecons = np.asarray(self.vaerecons, dtype=np.float32)
# self.states = np.asarray(self.states, dtype=np.float32)
self.update_reward(cnt, obs, action, reward)
# print(' '.join([('%.2f' % i)for i in self.rewards]))
# self.rewards = np.array(self.rewards)
# if os.path.exists('obs/data'):
# with open('obs/data', 'rb') as fp:
# obs, actions, values, neglogpacs, rewards, vaerecons, states = msgpack.load(fp, encoding='utf-8', raw=False)
# else:
# obs, actions, values, neglogpacs, rewards, vaerecons, states = [], [], [], [], [], [], []
# if len(obs) < MAX_SAVE:
# with open('obs/data', 'wb') as fp:
# msgpack.dump([obs + self.obs[l - BUFFER_LIMIT:l],\
# actions + self.actions[l - BUFFER_LIMIT:l],\
# values + self.values[l - BUFFER_LIMIT:l],\
# neglogpacs + self.neglogpacs[l - BUFFER_LIMIT:l],\
# rewards + self.rewards[l - BUFFER_LIMIT:l],\
# vaerecons + self.vaerecons[l - BUFFER_LIMIT:l],\
# states + self.states[l - BUFFER_LIMIT:l]],\
# fp, use_bin_type=True)
print(self.rewards[-20:])
print('Start Memory Replay')
self.memory_training()
print('Memory Replay Done')
def pre_process(self, inputs, refresh=False):
measurements, sensor_data, control, reward, steer, std_control, manual = inputs
sensor_data = self.process_sensor_data(sensor_data)
nowframe = np.concatenate([sensor_data[0], sensor_data[1]], 2)#深度和RGB图连接起来
if self.last_frame is None:
self.last_frame = nowframe
frame = np.concatenate([self.last_frame, nowframe], 2)#连续两帧连续起来
if refresh:
self.last_frame = nowframe
obs = (frame, measurements.player_measurements.forward_speed * 3.6 / 15.0 - 1, steer) #obs分为当前frame和speed
#将control从VehicleControl()变为数字
action = self.analyze_control(control)
std_action = self.analyze_control(std_control)
# if std_action == 0:
# manual = False
return obs, reward, action, std_action, manual
def update(self, inputs):
obs, reward, action, std_action, manual = self.pre_process(inputs, refresh=True)
# sensor_data = self.process_sensor_data(sensor_data)
# obs = [sensor_data[0]]
# auxs = [sensor_data[1], sensor_data[2],\
# abs(measurements.player_measurements.forward_speed) * 3.6 / 100, \
# collision / 20000,\
# measurements.player_measurements.intersection_offroad]
# control = self.analyze_control(control)
# reward = [reward]
# assert len(self.obs_buffer) == len(self.auxs_buffer) and len(self.auxs_buffer) == len(self.control_buffer)\
# and len(self.control_buffer) == len(self.reward_buffer)
self.states.append(self.state)
_, value, self.state, neglogpacs, vaerecon = self.machine.value(obs, self.state, action)
self.obs.append(obs)
self.actions.append(action)
self.values.append(value)
self.neglogpacs.append(neglogpacs)
self.rewards.append(reward)
self.vaerecons.append(vaerecon)
self.std_actions.append(std_action)
self.manual.append(manual)
# self.red_buffer.append(red)
# self.manual_buffer.append(manual)
def pretrain(self):
raise NotImplementedError
# if os.path.exists('obs/data'):
# print('Start Pretraining!!')
# with open('obs/data', 'rb') as fp:
# self.obs, self.actions, self.values, self.neglogpacs, self.rewards, self.vaerecons, self.states = msgpack.load(fp, encoding='utf-8', raw=False)
# print('Pretraining length = ', len(self.obs))
# self.memory_training(pretrain=True)
def calculate_difficulty(self, reward, vaerecon):
# return abs(reward)
return 1
def memory_training(self, pretrain=False):
l = len(self.obs)
batch = []
difficulty = []
for i in range(l):
batch.append([self.obs[i], self.actions[i], self.values[i], self.neglogpacs[i], self.rewards[i], self.vaerecons[i], self.states[i], self.std_actions[i], self.manual[i]])
difficulty.append(self.calculate_difficulty(self.rewards[i], self.vaerecons[i]))
# print(self.rewards)
# print(self.values)
# print(np.array(self.rewards) - np.array([i[0] for i in self.values]))
difficulty = np.array(difficulty)
print(difficulty[-20:])
def softmax(x):
x = np.clip(x, 1e-3, 1)
return np.exp(x) / np.sum(np.exp(x), axis=0)
difficulty = softmax(difficulty * 5)
print(difficulty[-20:])
print("Memory Extraction Done.")
for _ in tqdm(range(TRAIN_EPOCH)):
roll = np.random.choice(len(difficulty), BATCH_SIZE, p=difficulty)
tbatch = []
for i in roll:
tbatch.append(batch[i])
tra_batch = [np.array([t[i] for t in tbatch]) for i in range(9)]
# tra_batch = [np.array([t[i] for t in tbatch]) for i in range(7)]
self.machine.train(tra_batch, self.global_step)
self.global_step += 1
self.machine.save()
# if pretrain:
# self.obs, self.actions, self.values, self.neglogpacs, self.rewards, self.vaerecons, self.states = [],[],[],[],[],[],[]
if len(self.obs) > KEEP_CNT:
rem = len(self.obs) - KEEP_CNT
self.obs, self.actions, self.values, self.neglogpacs, self.rewards, self.vaerecons, self.states, self.std_actions, self.manual = \
self.obs[rem:],self.actions[rem:],self.values[rem:],self.neglogpacs[rem:],self.rewards[rem:],self.vaerecons[rem:], self.states[rem:], self.std_actions[rem:], self.manual[rem:]
def decode_control(self, cod):
#将数字的ｃｏｎｔｒｏｌ转换为VehicleControl()
control = VehicleControl()
control.steer = 0
control.throttle = 0
control.brake = 0
control.hand_brake = False
control.reverse = False
th, steer = cod
if th > 0:
control.throttle = min(th, 1.0)
if th < 0:
control.brake = min(-th, 1.0)
control.steer = max(min(steer, 1.0), -1.0)
# if cod > 9:
# control.hand_brake = True
# if cod == 10:
# control.steer = -1
# elif cod == 11:
# control.steer = 1
# return control
# if cod == 9:
# control.reverse = True
# control.throttle = 1
# return control
# if cod % 3 == 1:
# control.brake = 1
# elif cod % 3 == 2:
# control.throttle = 1
# if cod // 3 == 1:
# control.steer = 1
# elif cod // 3 == 2:
# control.steer = -1
return control
def get_control(self, inputs):
obs, reward, action, std_action, manual = self.pre_process(inputs)
print(action)
print(std_action)
action, _, _, _, _ = self.machine.step(obs, self.state)#整个模型跑一步
print(action)
control = self.decode_control(action[0])
return control