-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_3_actions.py
258 lines (231 loc) · 9.3 KB
/
train_3_actions.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
import torch
import gym
import numpy as np
from TD3 import TD3
from utils import ReplayBuffer
from collections import deque
import pickle
import matplotlib.pyplot as plt
from typing import Union
from gym import spaces
from gym.spaces import Box
######### Custom Gym Wrapper Function #########
class CustomWrapper(gym.Wrapper):
# def __init__(self, env):
# # super().__init__(env)
# low = self.action_space.low[:3]
# high = self.action_space.high[:3]
# self.action_space = spaces.Box(
# low,
# high,
# # shape=env.action_space.shape,
# dtype=env.action_space.dtype,
# )
#state space
# Num Observation Min Max Mean
# 0 hull_angle 0 2*pi 0.5
# 1 hull_angularVelocity -inf +inf -
# 2 vel_x -1 +1 -
# 3 vel_y -1 +1 -
# 4 hip_joint_1_angle -inf +inf -
# 5 hip_joint_1_speed -inf +inf -
# 6 knee_joint_1_angle -inf +inf -
# 7 knee_joint_1_speed -inf +inf -
# 8 leg_1_ground_contact_flag 0 1 -
# 9 hip_joint_2_angle -inf +inf -
# 10 hip_joint_2_speed -inf +inf -
# 11 knee_joint_2_angle -inf +inf -
# 12 knee_joint_2_speed -inf +inf -
# 13 leg_2_ground_contact_flag 0 1 -
# 14-23 10 lidar readings -inf +inf -
#action space
# Num Name Min Max
# 0 Hip_1 (Torque / Velocity) -1 +1
# 1 Knee_1 (Torque / Velocity) -1 +1
# 2 Hip_2 (Torque / Velocity) -1 +1
# 3 Knee_2 (Torque / Velocity) -1 +1
def __init__(
self,
env: gym.Env,
# min_action: Union[float, int, np.ndarray],
# max_action: Union[float, int, np.ndarray],
):
"""Initializes the :class:`RescaleAction` wrapper.
Args:
env (Env): The environment to apply the wrapper
min_action (float, int or np.ndarray): The min values for each action. This may be a numpy array or a scalar.
max_action (float, int or np.ndarray): The max values for each action. This may be a numpy array or a scalar.
"""
# assert isinstance(
# env.action_space, spaces.Box
# ), f"expected Box action space, got {type(env.action_space)}"
# assert np.less_equal(min_action, max_action).all(), (min_action, max_action)
super().__init__(env)
# self.min_action = (
# np.zeros(env.action_space.shape, dtype=env.action_space.dtype) + min_action
# )
# self.max_action = (
# np.zeros(env.action_space.shape, dtype=env.action_space.dtype) + max_action
# )
low = self.action_space.low[:3]
new_low = np.append(low,0)
high = self.action_space.high[:3]
new_high = np.append(high,0)
self.action_space = spaces.Box(
new_low,
new_high,
# shape=env.action_space.shape,
dtype=env.action_space.dtype,
)
# # low = np.append(self.observation_space.low, 0.0)
# low = self.observation_space.low[:19]
# # high = np.append(self.observation_space.high, np.inf)
# high = self.observation_space.high[:19]
# self.observation_space = Box(low, high, dtype=np.float32)
def step(self, action):
# modify obs
obs, reward, terminated, info = self.env.step(action)
# obs = obs[:19]
return obs, reward, terminated, info
def reset(self):
obs = self.env.reset()
# obs = obs[:19]
return obs
# def action(self, action):
# """Rescales the action affinely from [:attr:`min_action`, :attr:`max_action`] to the action space of the base environment, :attr:`env`.
# Args:
# action: The action to rescale
# Returns:
# The rescaled action
# """
# assert np.all(np.greater_equal(action, self.min_action)), (
# action,
# self.min_action,
# )
# assert np.all(np.less_equal(action, self.max_action)), (action, self.max_action)
# low = self.env.action_space.low
# high = self.env.action_space.high
# action = low + (high - low) * (
# (action - self.min_action) / (self.max_action - self.min_action)
# )
# action = np.clip(action, low, high)
# return action
def action(self,action):
reduced_action = action[:3]
new_action = np.append(reduced_action,0)
return new_action
######### Hyperparameters #########
gym.logger.set_level(40)
env_name = "BipedalWalker-v3"
env = CustomWrapper(gym.make(env_name))
# env = gym.make(env_name)
state_dim = env.observation_space.shape[0]
print('state dimensions', state_dim)
# state_dim = 20
action_dim = env.action_space.shape[0]
print('action space', action_dim)
max_action = float(env.action_space.high[0])
# print('obs space ',env.observation_space.shape[0])
# print('action space high',env.action_space.high)
log_interval = 100 # print avg reward after interval
random_seed = 0
gamma = 0.99 # discount for future rewards
batch_size = 100 # num of transitions sampled from replay buffer
lr = 0.001
exploration_noise = 0.1
polyak = 0.995 # target policy update parameter (1-tau)
policy_noise = 0.2 # target policy smoothing noise
noise_clip = 0.3
policy_delay = 2 # delayed policy updates parameter
max_episodes = 10000 # max num of episodes
max_timesteps = 2000 # max timesteps in one episode
directory = "./preTrained_3_action/" # save trained models
filename = "TD3_{}_{}".format(env_name, random_seed)
start_episode = 0
print(state_dim)
policy = TD3(lr, state_dim, action_dim, max_action)
replay_buffer = ReplayBuffer()
if random_seed:
print("Random Seed: {}".format(random_seed))
env.seed(random_seed)
torch.manual_seed(random_seed)
np.random.seed(random_seed)
LOAD = False
if LOAD:
start_episode = 6
policy.load(directory, filename, str(start_episode))
# logging variables:
scores = []
mean_scores = []
last_scores = deque(maxlen=log_interval)
distances = []
mean_distances = []
last_distance = deque(maxlen=log_interval)
losses_mean_episode = []
prune_ct = 200
# training procedure:
for ep in range(start_episode + 1, max_episodes + 1):
state = env.reset()
total_reward = 0
total_distance = 0
actor_losses = []
c1_losses = []
c2_losses = []
# if ep % prune_ct == 0:
# policy.prune_model()
# print('i have pruned your model by 10%')
for t in range(max_timesteps):
# select action and add exploration noise:
action = policy.select_action(state)
# print('action size',action.shape)
action = action + np.random.normal(0, exploration_noise, size=env.action_space.shape[0])
action = action.clip(env.action_space.low, env.action_space.high)
# take action in env:
next_state, reward, done, _ = env.step(action)
replay_buffer.add((state, action, reward, next_state, float(done)))
state = next_state
# print('state size', state.shape)
total_reward += reward
if reward != -100:
total_distance += reward
# if episode is done then update policy:
if done or t == (max_timesteps - 1):
actor_loss, c1_loss, c2_loss = policy.update(replay_buffer, t, batch_size, gamma, polyak, policy_noise, noise_clip, policy_delay)
actor_losses.append(actor_loss)
c1_losses.append((c1_loss))
c2_losses.append(c2_loss)
break
mean_loss_actor = np.mean(actor_losses)
mean_loss_c1 = np.mean(c1_losses)
mean_loss_c2 = np.mean(c2_losses)
losses_mean_episode.append((ep, mean_loss_actor, mean_loss_c1, mean_loss_c2))
print('\rEpisode: {}/{},\tScore: {:.2f},\tDistance: {:.2f},\tactor_loss: {},\tc1_loss:{},\tc2_loss:{}'.format(ep, max_episodes,total_reward,total_distance,mean_loss_actor,mean_loss_c1, mean_loss_c2))
# logging updates:
scores.append(total_reward)
distances.append(total_distance)
last_scores.append(total_reward)
last_distance.append(total_distance)
mean_score = np.mean(last_scores)
mean_distance = np.mean(last_distance)
FILE = 'record_3_action.dat'
data = [ep, total_reward, total_distance, mean_loss_actor, mean_loss_c1, mean_loss_c2]
with open(FILE, "ab") as f:
pickle.dump(data, f)
# if avg reward > 300 then save and stop traning:
if (mean_score) >= 300:
print("########## Solved! ###########")
name = filename + '_solved'
policy.save(directory, name, str(ep))
break
# print avg reward every log interval:
if ep % log_interval == 0:
policy.save(directory, filename, str(ep))
mean_scores.append(mean_score)
mean_distances.append(mean_distance)
print('\rEpisode: {}/{},\tMean Score: {:.2f},\tMean Distance: {:.2f},\tactor_loss: {},\tc1_loss:{},\tc2_loss:{}'
.format(ep, max_episodes, mean_score, mean_distance, mean_loss_actor, mean_loss_c1, mean_loss_c2))
FILE = 'record_3_action_mean.dat'
data = [ep, mean_score, mean_distance, mean_loss_actor, mean_loss_c1, mean_loss_c2]
with open(FILE, "ab") as f:
pickle.dump(data, f)
env.close()