-
Notifications
You must be signed in to change notification settings - Fork 129
/
sac_discrete_per.py
305 lines (238 loc) · 12 KB
/
sac_discrete_per.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
'''
Soft Actor-Critic version 2
using target Q instead of V net: 2 Q net, 2 target Q net, 1 policy net
add alpha loss compared with version 1
paper: https://arxiv.org/pdf/1812.05905.pdf
Discrete version reference:
https://towardsdatascience.com/adapting-soft-actor-critic-for-discrete-action-spaces-a20614d4a50a
'''
import random
import gym
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions import Categorical
from IPython.display import clear_output
import matplotlib.pyplot as plt
import argparse
from common.buffers import ReplayBufferPER
GPU = True
device_idx = 0
if GPU:
device = torch.device("cuda:" + str(device_idx) if torch.cuda.is_available() else "cpu")
else:
device = torch.device("cpu")
print(device)
parser = argparse.ArgumentParser(description='Train or test neural net motor controller.')
parser.add_argument('--train', dest='train', action='store_true', default=False)
parser.add_argument('--test', dest='test', action='store_true', default=False)
args = parser.parse_args()
class SoftQNetwork(nn.Module):
def __init__(self, num_inputs, num_actions, hidden_size, init_w=3e-3):
super(SoftQNetwork, self).__init__()
self.linear1 = nn.Linear(num_inputs, hidden_size)
self.linear2 = nn.Linear(hidden_size, hidden_size)
# self.linear3 = nn.Linear(hidden_size, hidden_size)
self.linear4 = nn.Linear(hidden_size, num_actions)
self.linear4.weight.data.uniform_(-init_w, init_w)
self.linear4.bias.data.uniform_(-init_w, init_w)
def forward(self, state):
x = F.tanh(self.linear1(state))
x = F.tanh(self.linear2(x))
# x = F.tanh(self.linear3(x))
x = self.linear4(x)
return x
class PolicyNetwork(nn.Module):
def __init__(self, num_inputs, num_actions, hidden_size, init_w=3e-3, log_std_min=-20, log_std_max=2):
super(PolicyNetwork, self).__init__()
self.linear1 = nn.Linear(num_inputs, hidden_size)
self.linear2 = nn.Linear(hidden_size, hidden_size)
# self.linear3 = nn.Linear(hidden_size, hidden_size)
# self.linear4 = nn.Linear(hidden_size, hidden_size)
self.output = nn.Linear(hidden_size, num_actions)
self.num_actions = num_actions
def forward(self, state, softmax_dim=-1):
x = F.tanh(self.linear1(state))
x = F.tanh(self.linear2(x))
# x = F.tanh(self.linear3(x))
# x = F.tanh(self.linear4(x))
probs = F.softmax(self.output(x), dim=softmax_dim)
return probs
def evaluate(self, state, epsilon=1e-8):
'''
generate sampled action with state as input wrt the policy network;
'''
probs = self.forward(state, softmax_dim=-1)
log_probs = torch.log(probs)
# Avoid numerical instability. Ref: https://github.com/ku2482/sac-discrete.pytorch/blob/40c9d246621e658750e0a03001325006da57f2d4/sacd/model.py#L98
z = (probs == 0.0).float() * epsilon
log_probs = torch.log(probs + z)
return log_probs
def get_action(self, state, deterministic):
state = torch.FloatTensor(state).unsqueeze(0).to(device)
probs = self.forward(state)
dist = Categorical(probs)
if deterministic:
action = np.argmax(probs.detach().cpu().numpy())
else:
action = dist.sample().squeeze().detach().cpu().numpy()
return action
class SAC_Trainer():
def __init__(self, replay_buffer, hidden_dim):
self.replay_buffer = replay_buffer
self.soft_q_net1 = SoftQNetwork(state_dim, action_dim, hidden_dim).to(device)
self.soft_q_net2 = SoftQNetwork(state_dim, action_dim, hidden_dim).to(device)
self.target_soft_q_net1 = SoftQNetwork(state_dim, action_dim, hidden_dim).to(device)
self.target_soft_q_net2 = SoftQNetwork(state_dim, action_dim, hidden_dim).to(device)
self.policy_net = PolicyNetwork(state_dim, action_dim, hidden_dim).to(device)
self.log_alpha = torch.zeros(1, dtype=torch.float32, requires_grad=True, device=device)
print('Soft Q Network (1,2): ', self.soft_q_net1)
print('Policy Network: ', self.policy_net)
for target_param, param in zip(self.target_soft_q_net1.parameters(), self.soft_q_net1.parameters()):
target_param.data.copy_(param.data)
for target_param, param in zip(self.target_soft_q_net2.parameters(), self.soft_q_net2.parameters()):
target_param.data.copy_(param.data)
self.soft_q_criterion1 = nn.MSELoss(reduction="none")
self.soft_q_criterion2 = nn.MSELoss(reduction="none")
soft_q_lr = 3e-4
policy_lr = 3e-4
alpha_lr = 3e-4
self.soft_q_optimizer1 = optim.Adam(self.soft_q_net1.parameters(), lr=soft_q_lr)
self.soft_q_optimizer2 = optim.Adam(self.soft_q_net2.parameters(), lr=soft_q_lr)
self.policy_optimizer = optim.Adam(self.policy_net.parameters(), lr=policy_lr)
self.alpha_optimizer = optim.Adam([self.log_alpha], lr=alpha_lr)
def update(self, batch_size, reward_scale=10., auto_entropy=True, target_entropy=-2, gamma=0.99, soft_tau=1e-2):
state, action, reward, next_state, done = self.replay_buffer.sample(batch_size)
# print('sample:', state, action, reward, done)
state = torch.FloatTensor(state).to(device)
next_state = torch.FloatTensor(next_state).to(device)
action = torch.Tensor(action).to(torch.int64).to(device)
reward = torch.FloatTensor(reward).unsqueeze(1).to(device) # reward is single value, unsqueeze() to add one dim to be [reward] at the sample dim;
done = torch.FloatTensor(np.float32(done)).unsqueeze(1).to(device)
predicted_q_value1 = self.soft_q_net1(state)
predicted_q_value1 = predicted_q_value1.gather(1, action.unsqueeze(-1))
predicted_q_value2 = self.soft_q_net2(state)
predicted_q_value2 = predicted_q_value2.gather(1, action.unsqueeze(-1))
log_prob = self.policy_net.evaluate(state)
next_log_prob = self.policy_net.evaluate(next_state)
# reward = reward_scale * (reward - reward.mean(dim=0)) / (reward.std(dim=0) + 1e-6) # normalize with batch mean and std; plus a small number to prevent numerical problem
# Training Q Function
self.alpha = self.log_alpha.exp()
target_q_min = (next_log_prob.exp() * (torch.min(self.target_soft_q_net1(next_state),self.target_soft_q_net2(next_state)) - self.alpha * next_log_prob)).sum(dim=-1).unsqueeze(-1)
target_q_value = reward + (1 - done) * gamma * target_q_min # if done==1, only reward
q_value_loss1 = self.soft_q_criterion1(predicted_q_value1, target_q_value.detach()) # detach: no gradients for the variable
q_value_loss2 = self.soft_q_criterion2(predicted_q_value2, target_q_value.detach())
weight_update = [min(l1.item(), l2.item()) for l1, l2 in zip(q_value_loss1, q_value_loss2)]
self.replay_buffer.update_weights(weight_update) # update sample weights with td error
self.soft_q_optimizer1.zero_grad()
q_value_loss1.mean().backward()
self.soft_q_optimizer1.step()
self.soft_q_optimizer2.zero_grad()
q_value_loss2.mean().backward()
self.soft_q_optimizer2.step()
# Training Policy Function
predicted_new_q_value = torch.min(self.soft_q_net1(state),self.soft_q_net2(state))
policy_loss = (log_prob.exp()*(self.alpha * log_prob - predicted_new_q_value)).sum(dim=-1).mean()
self.policy_optimizer.zero_grad()
policy_loss.backward()
self.policy_optimizer.step()
# Updating alpha wrt entropy
# alpha = 0.0 # trade-off between exploration (max entropy) and exploitation (max Q)
if auto_entropy is True:
alpha_loss = -(self.log_alpha * (log_prob + target_entropy).detach()).mean()
# print('alpha loss: ',alpha_loss)
self.alpha_optimizer.zero_grad()
alpha_loss.backward()
self.alpha_optimizer.step()
else:
self.alpha = 1.
alpha_loss = 0
# print('q loss: ', q_value_loss1.item(), q_value_loss2.item())
# print('policy loss: ', policy_loss.item() )
# Soft update the target value net
for target_param, param in zip(self.target_soft_q_net1.parameters(), self.soft_q_net1.parameters()):
target_param.data.copy_( # copy data value into target parameters
target_param.data * (1.0 - soft_tau) + param.data * soft_tau
)
for target_param, param in zip(self.target_soft_q_net2.parameters(), self.soft_q_net2.parameters()):
target_param.data.copy_( # copy data value into target parameters
target_param.data * (1.0 - soft_tau) + param.data * soft_tau
)
return predicted_new_q_value.mean()
def save_model(self, path):
torch.save(self.soft_q_net1.state_dict(), path+'_q1')
torch.save(self.soft_q_net2.state_dict(), path+'_q2')
torch.save(self.policy_net.state_dict(), path+'_policy')
def load_model(self, path):
self.soft_q_net1.load_state_dict(torch.load(path+'_q1'))
self.soft_q_net2.load_state_dict(torch.load(path+'_q2'))
self.policy_net.load_state_dict(torch.load(path+'_policy'))
self.soft_q_net1.eval()
self.soft_q_net2.eval()
self.policy_net.eval()
def plot(rewards):
clear_output(True)
plt.figure(figsize=(20,5))
plt.plot(rewards)
plt.savefig('sac_v2.png')
# plt.show()
replay_buffer_size = 1e6
replay_buffer = ReplayBufferPER(replay_buffer_size)
# choose env
env = gym.make('CartPole-v1')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.n # discrete
# hyper-parameters for RL training
max_episodes = 10000
max_steps = 200
frame_idx = 0
batch_size = 256
update_itr = 1
AUTO_ENTROPY=True
DETERMINISTIC=False
hidden_dim = 64
rewards = []
model_path = './model/sac_discrete_v2'
target_entropy = -1.*action_dim
# target_entropy = 0.98 * -np.log(1 / action_dim)
sac_trainer=SAC_Trainer(replay_buffer, hidden_dim=hidden_dim)
if __name__ == '__main__':
if args.train:
# training loop
for eps in range(max_episodes):
state = env.reset()
episode_reward = 0
for step in range(max_steps):
action = sac_trainer.policy_net.get_action(state, deterministic = DETERMINISTIC)
next_state, reward, done, _ = env.step(action)
# env.render()
replay_buffer.push(state, action, reward, next_state, done)
state = next_state
episode_reward += reward
frame_idx += 1
if len(replay_buffer) > batch_size:
for i in range(update_itr):
_=sac_trainer.update(batch_size, reward_scale=1., auto_entropy=AUTO_ENTROPY, target_entropy=target_entropy)
if done:
break
if eps % 20 == 0 and eps>0: # plot and model saving interval
plot(rewards)
np.save('rewards', rewards)
sac_trainer.save_model(model_path)
print('Episode: ', eps, '| Episode Reward: ', episode_reward, '| Episode Length: ', step)
rewards.append(episode_reward)
sac_trainer.save_model(model_path)
if args.test:
sac_trainer.load_model(model_path)
for eps in range(10):
state = env.reset()
episode_reward = 0
for step in range(max_steps):
action = sac_trainer.policy_net.get_action(state, deterministic = DETERMINISTIC)
next_state, reward, done, _ = env.step(action)
env.render()
episode_reward += reward
state=next_state
print('Episode: ', eps, '| Episode Reward: ', episode_reward)