/
ppo_agent.py
213 lines (208 loc) · 10.7 KB
/
ppo_agent.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
import numpy as np
import torch
from torch import optim
from running_state import ZFilter
from utils import select_actions, evaluate_actions
from datetime import datetime
import os
import copy
class ppo_agent:
def __init__(self, envs, args, net, env_type='atari'):
self.envs = envs
self.args = args
self.env_type = env_type
# define the newtork...
self.net = net
self.old_net = copy.deepcopy(self.net)
# if use the cuda...
if self.args.cuda:
self.net.cuda()
self.old_net.cuda()
# define the optimizer...
self.optimizer = optim.Adam(self.net.parameters(), self.args.lr, eps=self.args.eps)
# running filter...
if self.env_type == 'mujoco':
num_states = self.envs.observation_space.shape[0]
self.running_state = ZFilter((num_states, ), clip=5)
# check saving folder..
if not os.path.exists(self.args.save_dir):
os.mkdir(self.args.save_dir)
# env folder..
self.model_path = os.path.join(self.args.save_dir, self.args.env_name)
if not os.path.exists(self.model_path):
os.mkdir(self.model_path)
# get the observation
self.batch_ob_shape = (self.args.num_workers * self.args.nsteps, ) + self.envs.observation_space.shape
self.obs = np.zeros((self.args.num_workers, ) + self.envs.observation_space.shape, dtype=self.envs.observation_space.dtype.name)
if self.env_type == 'mujoco':
self.obs[:] = np.expand_dims(self.running_state(self.envs.reset()), 0)
else:
self.obs[:] = self.envs.reset()
self.dones = [False for _ in range(self.args.num_workers)]
# start to train the network...
def learn(self):
num_updates = self.args.total_frames // (self.args.nsteps * self.args.num_workers)
# get the reward to calculate other informations
episode_rewards = torch.zeros([self.args.num_workers, 1])
final_rewards = torch.zeros([self.args.num_workers, 1])
for update in range(num_updates):
mb_obs, mb_rewards, mb_actions, mb_dones, mb_values = [], [], [], [], []
if self.args.lr_decay:
self._adjust_learning_rate(update, num_updates)
for step in range(self.args.nsteps):
with torch.no_grad():
# get tensors
obs_tensor = self._get_tensors(self.obs)
values, pis = self.net(obs_tensor)
# select actions
actions = select_actions(pis, self.args.dist, self.env_type)
if self.env_type == 'atari':
input_actions = actions
else:
if self.args.dist == 'gauss':
input_actions = actions.copy()
elif self.args.dist == 'beta':
input_actions = -1 + 2 * actions
# start to store information
mb_obs.append(np.copy(self.obs))
mb_actions.append(actions)
mb_dones.append(self.dones)
mb_values.append(values.detach().cpu().numpy().squeeze())
# start to excute the actions in the environment
obs, rewards, dones, _ = self.envs.step(input_actions)
# update dones
if self.env_type == 'mujoco':
dones = np.array([dones])
rewards = np.array([rewards])
self.dones = dones
mb_rewards.append(rewards)
# clear the observation
for n, done in enumerate(dones):
if done:
self.obs[n] = self.obs[n] * 0
if self.env_type == 'mujoco':
# reset the environment
obs = self.envs.reset()
self.obs = obs if self.env_type == 'atari' else np.expand_dims(self.running_state(obs), 0)
# process the rewards part -- display the rewards on the screen
rewards = torch.tensor(np.expand_dims(np.stack(rewards), 1), dtype=torch.float32)
episode_rewards += rewards
masks = torch.tensor([[0.0] if done_ else [1.0] for done_ in dones], dtype=torch.float32)
final_rewards *= masks
final_rewards += (1 - masks) * episode_rewards
episode_rewards *= masks
# process the rollouts
mb_obs = np.asarray(mb_obs, dtype=np.float32)
mb_rewards = np.asarray(mb_rewards, dtype=np.float32)
mb_actions = np.asarray(mb_actions, dtype=np.float32)
mb_dones = np.asarray(mb_dones, dtype=np.bool)
mb_values = np.asarray(mb_values, dtype=np.float32)
if self.env_type == 'mujoco':
mb_values = np.expand_dims(mb_values, 1)
# compute the last state value
with torch.no_grad():
obs_tensor = self._get_tensors(self.obs)
last_values, _ = self.net(obs_tensor)
last_values = last_values.detach().cpu().numpy().squeeze()
# start to compute advantages...
mb_returns = np.zeros_like(mb_rewards)
mb_advs = np.zeros_like(mb_rewards)
lastgaelam = 0
for t in reversed(range(self.args.nsteps)):
if t == self.args.nsteps - 1:
nextnonterminal = 1.0 - self.dones
nextvalues = last_values
else:
nextnonterminal = 1.0 - mb_dones[t + 1]
nextvalues = mb_values[t + 1]
delta = mb_rewards[t] + self.args.gamma * nextvalues * nextnonterminal - mb_values[t]
mb_advs[t] = lastgaelam = delta + self.args.gamma * self.args.tau * nextnonterminal * lastgaelam
mb_returns = mb_advs + mb_values
# after compute the returns, let's process the rollouts
mb_obs = mb_obs.swapaxes(0, 1).reshape(self.batch_ob_shape)
if self.env_type == 'atari':
mb_actions = mb_actions.swapaxes(0, 1).flatten()
mb_returns = mb_returns.swapaxes(0, 1).flatten()
mb_advs = mb_advs.swapaxes(0, 1).flatten()
# before update the network, the old network will try to load the weights
self.old_net.load_state_dict(self.net.state_dict())
# start to update the network
pl, vl, ent = self._update_network(mb_obs, mb_actions, mb_returns, mb_advs)
# display the training information
if update % self.args.display_interval == 0:
print('[{}] Update: {} / {}, Frames: {}, Rewards: {:.3f}, Min: {:.3f}, Max: {:.3f}, PL: {:.3f},'\
'VL: {:.3f}, Ent: {:.3f}'.format(datetime.now(), update, num_updates, (update + 1)*self.args.nsteps*self.args.num_workers, \
final_rewards.mean().item(), final_rewards.min().item(), final_rewards.max().item(), pl, vl, ent))
# save the model
if self.env_type == 'atari':
torch.save(self.net.state_dict(), self.model_path + '/model.pt')
else:
# for the mujoco, we also need to keep the running mean filter!
torch.save([self.net.state_dict(), self.running_state], self.model_path + '/model.pt')
# update the network
def _update_network(self, obs, actions, returns, advantages):
inds = np.arange(obs.shape[0])
nbatch_train = obs.shape[0] // self.args.batch_size
for _ in range(self.args.epoch):
np.random.shuffle(inds)
for start in range(0, obs.shape[0], nbatch_train):
# get the mini-batchs
end = start + nbatch_train
mbinds = inds[start:end]
mb_obs = obs[mbinds]
mb_actions = actions[mbinds]
mb_returns = returns[mbinds]
mb_advs = advantages[mbinds]
# convert minibatches to tensor
mb_obs = self._get_tensors(mb_obs)
mb_actions = torch.tensor(mb_actions, dtype=torch.float32)
mb_returns = torch.tensor(mb_returns, dtype=torch.float32).unsqueeze(1)
mb_advs = torch.tensor(mb_advs, dtype=torch.float32).unsqueeze(1)
# normalize adv
mb_advs = (mb_advs - mb_advs.mean()) / (mb_advs.std() + 1e-8)
if self.args.cuda:
mb_actions = mb_actions.cuda()
mb_returns = mb_returns.cuda()
mb_advs = mb_advs.cuda()
# start to get values
mb_values, pis = self.net(mb_obs)
# start to calculate the value loss...
value_loss = (mb_returns - mb_values).pow(2).mean()
# start to calculate the policy loss
with torch.no_grad():
_, old_pis = self.old_net(mb_obs)
# get the old log probs
old_log_prob, _ = evaluate_actions(old_pis, mb_actions, self.args.dist, self.env_type)
old_log_prob = old_log_prob.detach()
# evaluate the current policy
log_prob, ent_loss = evaluate_actions(pis, mb_actions, self.args.dist, self.env_type)
prob_ratio = torch.exp(log_prob - old_log_prob)
# surr1
surr1 = prob_ratio * mb_advs
surr2 = torch.clamp(prob_ratio, 1 - self.args.clip, 1 + self.args.clip) * mb_advs
policy_loss = -torch.min(surr1, surr2).mean()
# final total loss
total_loss = policy_loss + self.args.vloss_coef * value_loss - ent_loss * self.args.ent_coef
# clear the grad buffer
self.optimizer.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(self.net.parameters(), self.args.max_grad_norm)
# update
self.optimizer.step()
return policy_loss.item(), value_loss.item(), ent_loss.item()
# convert the numpy array to tensors
def _get_tensors(self, obs):
if self.env_type == 'atari':
obs_tensor = torch.tensor(np.transpose(obs, (0, 3, 1, 2)), dtype=torch.float32)
else:
obs_tensor = torch.tensor(obs, dtype=torch.float32)
# decide if put the tensor on the GPU
if self.args.cuda:
obs_tensor = obs_tensor.cuda()
return obs_tensor
# adjust the learning rate
def _adjust_learning_rate(self, update, num_updates):
lr_frac = 1 - (update / num_updates)
adjust_lr = self.args.lr * lr_frac
for param_group in self.optimizer.param_groups:
param_group['lr'] = adjust_lr