-
Notifications
You must be signed in to change notification settings - Fork 0
/
PPO.py
212 lines (174 loc) · 7.6 KB
/
PPO.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
import torch
import torch.nn as nn
from torch.distributions import Categorical
import gym
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class Memory():
def __init__(self):
self.actions = []
self.states = []
self.logprobs = []
self.rewards = []
self.is_terminals = []
def clear_memory(self):
del self.actions[:]
del self.states[:]
del self.logprobs[:]
del self.rewards[:]
del self.is_terminals[:]
class ActorCritic(nn.Module):
def __init__(self, state_dim, action_dim, n_latent_var):
super(ActorCritic, self).__init__()
# actor
self.action_layer = nn.Sequential(
nn.Linear(state_dim, n_latent_var),
nn.Tanh(),
nn.Linear(n_latent_var, n_latent_var),
nn.Tanh(),
nn.Linear(n_latent_var, action_dim),
nn.Softmax(dim=-1)
)
# critic
self.value_layer = nn.Sequential(
nn.Linear(state_dim, n_latent_var),
nn.Tanh(),
nn.Linear(n_latent_var, n_latent_var),
nn.Tanh(),
nn.Linear(n_latent_var, 1)
)
def forward(self):
raise NotImplementedError
# for interacting with environment
def act(self, state, memory):
state = torch.from_numpy(state).float().to(device)
action_probs = self.action_layer(state)
dist = Categorical(action_probs)
action = dist.sample()
memory.states.append(state)
memory.actions.append(action)
memory.logprobs.append(dist.log_prob(action))
return action.item()
# for ppo update
def evaluate(self, state, action):
action_probs = self.action_layer(state)
dist = Categorical(action_probs)
action_logprobs = dist.log_prob(action)
dist_entropy = dist.entropy()
state_value = self.value_layer(state)
# action_logprobs indirectly represents the policy $\pi_{\theta}(s,a)$
return action_logprobs, torch.squeeze(state_value), dist_entropy
class PPO():
def __init__(self, state_dim, action_dim, n_latent_var, lr, betas, gamma, K_epochs, eps_clip):
self.lr = lr
self.betas = betas
self.gamma = gamma
self.eps_clip = eps_clip
self.K_epochs = K_epochs
self.policy = ActorCritic(state_dim, action_dim, n_latent_var).to(device)
self.optimizer = torch.optim.Adam(self.policy.parameters(), lr=lr, betas=betas)
self.policy_old = ActorCritic(state_dim, action_dim, n_latent_var).to(device)
self.policy_old.load_state_dict(self.policy.state_dict())
self.MseLoss = nn.MSELoss()
def update(self, memory):
# Monte Carlo estimate of state rewards (can be replaced by General Advantage Estimators)
rewards = []
discounted_reward = 0
for reward, is_terminal in zip(reversed(memory.rewards), reversed(memory.is_terminals)):
if is_terminal:
discounted_reward = 0
discounted_reward = reward + (self.gamma * discounted_reward)
rewards.insert(0, discounted_reward)
# Normalizing the rewards
rewards = torch.tensor(rewards, dtype=torch.float32).to(device)
rewards = (rewards - rewards.mean()) / (rewards.std() + 1e-5)
# convert list to tensor
old_states = torch.stack(memory.states).to(device).detach()
old_actions = torch.stack(memory.actions).to(device).detach()
old_logprobs = torch.stack(memory.logprobs).to(device).detach()
# Optimize policy for K epochs
for _ in range(self.K_epochs):
# Evaluating old actions and values
logprobs, state_values, dist_entropy = self.policy.evaluate(old_states, old_actions)
# Finding the ratio (pi_theta / pi_theta__old)
ratios = torch.exp(logprobs - old_logprobs.detach())
# Finding Surrogate Loss (no gradient in advantages)
advantages = rewards - state_values.detach()
surr1 = ratios * advantages
surr2 = torch.clamp(ratios, 1-self.eps_clip, 1+self.eps_clip) * advantages
# MseLoss is for the update of critic, dist_entropy denotes an entropy bonus
loss = -torch.min(surr1, surr2) + 0.5*self.MseLoss(state_values, rewards) - 0.01*dist_entropy
# take gradient step
self.optimizer.zero_grad()
loss.mean().backward()
self.optimizer.step()
# Copy new weights into old policy:
self.policy_old.load_state_dict(self.policy.state_dict())
def main():
############## Hyperparameters ##############
env_name = "LunarLander-v2"
# creating environment
env = gym.make(env_name)
state_dim = env.observation_space.shape[0]
action_dim = 4
render = False
solved_reward = 230 # stop training if avg_reward > solved_reward
log_interval = 20 # print avg reward in the interval
max_episodes = 50000 # max training episodes
max_timesteps = 300 # max timesteps in one episode
n_latent_var = 64 # number of variables in hidden layer
update_timestep = 2000 # update policy every n timesteps
lr = 0.002
betas = (0.9, 0.999)
gamma = 0.99 # discount factor
K_epochs = 4 # update policy for K epochs
eps_clip = 0.2 # clip parameter for PPO
random_seed = None
#############################################
if random_seed:
torch.manual_seed(random_seed)
env.seed(random_seed)
memory = Memory()
ppo = PPO(state_dim, action_dim, n_latent_var, lr, betas, gamma, K_epochs, eps_clip)
print('learning rate:',lr, 'Adam betas:', betas)
# logging variables
running_reward = 0
avg_length = 0
timestep = 0
# training loop
for i_episode in range(1, max_episodes+1):
state = env.reset()
for t in range(max_timesteps):
timestep += 1
# Running policy_old:
# also append state, action, action_logprobs to the memory
with torch.no_grad():
action = ppo.policy_old.act(state, memory)
state, reward, done, _ = env.step(action)
# Saving reward and is_terminal:
memory.rewards.append(reward)
memory.is_terminals.append(done)
# update if its time
if timestep % update_timestep == 0:
ppo.update(memory)
memory.clear_memory()
timestep = 0
running_reward += reward
if render:
env.render()
if done:
break
avg_length += t
# stop training if avg_reward > solved_reward
if running_reward > (log_interval*solved_reward):
print("########## Solved! ##########")
torch.save(ppo.policy.state_dict(), './PPO_{}.pth'.format(env_name))
break
# logging
if i_episode % log_interval == 0:
avg_length = int(avg_length/log_interval)
running_reward = int((running_reward/log_interval))
print('Episode {} \t avg length: {} \t reward: {}'.format(i_episode, avg_length, running_reward))
running_reward = 0
avg_length = 0
if __name__ == '__main__':
main()