-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
executable file
·164 lines (133 loc) · 6.11 KB
/
train.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
from common.vec_env.vec_logger import VecLogger
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.distributions import Categorical
GAMMA = 0.99
TAU = 0.95
N_STEPS = 5
NUM_EPOCH = 4
NUM_MINIBATCH = 4
CLIP_GRAD = 0.5
PPO_EPS = 0.2
COEF_VALUE = 0.5
COEF_ENTROPY = 0.01
def train(args, venv, model, path, device):
N = args.num_processes
T = N_STEPS
K = NUM_EPOCH
M = NUM_MINIBATCH
assert N % M == 0
net = model(venv.observation_space.shape[0], venv.action_space.n).to(device)
net.train()
optimizer = optim.Adam(net.parameters(), lr=args.lr, amsgrad=args.amsgrad)
vlogger = VecLogger(N=N, path=path)
vlogger.add_model(net)
state = venv.reset()
hx_v = torch.zeros((N, 512)).to(device)
cx_v = torch.zeros((N, 512)).to(device)
t = 0
while t < args.num_timesteps:
# Run policy pi_old in environment for T timesteps
with torch.no_grad():
# use numpy while running, ppo1 is an off-policy method as acer
states = []
hxs = []
cxs = []
actions = []
rewards = []
dones = []
values = []
Rs = []
gaes = []
log_prob_actions = []
# no gradient, otherwise GPU memory will leak
for step in range(T):
# save state, hx, cx before perform
states.append(state.copy())
hx = hx_v.data.cpu().numpy()
cx = cx_v.data.cpu().numpy()
hxs.append(hx.copy())
cxs.append(cx.copy())
# perform action according to policy
state_v = torch.from_numpy(state).float().to(device)
value_v, logit_v, (hx_v, cx_v) = net(state_v, (hx_v, cx_v))
dist_v = Categorical(logits=logit_v)
action_v = dist_v.sample()
log_prob_action_v = dist_v.log_prob(action_v)
action_v = action_v.unsqueeze(1)
log_prob_action_v = log_prob_action_v.unsqueeze(1)
# receive reward and new state
action = action_v.data.cpu().numpy()
state, reward, done, info = venv.step(action)
t += N
reward = np.expand_dims(reward, axis=1)
done = np.expand_dims(done, axis=1)
info = np.expand_dims(info, axis=1)
vlogger.log(t, reward, info)
value = value_v.data.cpu().numpy()
log_prob_action = log_prob_action_v.data.cpu().numpy()
actions.append(action.copy())
rewards.append(reward.copy())
dones.append(done.copy())
values.append(value.copy())
log_prob_actions.append(log_prob_action.copy())
# reset the LSTM state if done
done_v = torch.from_numpy(done.astype('int')).float().to(device)
hx_v = (1 - done_v) * hx_v
cx_v = (1 - done_v) * cx_v
# last value
state_v = torch.from_numpy(state).float().to(device)
value_v = net(state_v, (hx_v, cx_v))[0]
value = value_v.data.cpu().numpy()
values.append(value.copy())
# Compute advantage estimates
R = (1 - done) * value
gae = np.zeros((N, 1))
for i in reversed(range(T)):
# reference values
R = (1 - dones[i]) * GAMMA * R + rewards[i]
Rs.insert(0, R.copy())
# generalized advantage estimataion
delta_t = rewards[i] + (1 - dones[i]) * GAMMA * values[i + 1] - values[i]
gae = gae * (1 - dones[i]) * GAMMA * TAU + delta_t
gaes.insert(0, gae.copy())
states = np.array(states)
hxs = np.array(hxs)
cxs = np.array(cxs)
actions = np.array(actions)
Rs = np.array(Rs)
gaes = np.array(gaes)
log_prob_actions = np.array(log_prob_actions)
# gaes = (gaes - gaes.mean()) / (gaes.std() + 1e-8)
# Optimize surrogate L wrt theta, with K epochs and minibatch size NM <= NT
for _ in range(K):
indices = np.random.permutation(N)
for k in range(0, N, M):
# minibatch tensors shape (T*M, ...), main bottleneck for GPU memory
state_vs = torch.from_numpy(states[:, indices[k:k+M], ...].reshape(T*M, 1, 84, 84)).float().to(device)
hx_vs = torch.from_numpy(hxs[:, indices[k:k+M], ...].reshape(T*M, -1)).float().to(device)
cx_vs = torch.from_numpy(cxs[:, indices[k:k+M], ...].reshape(T*M, -1)).float().to(device)
action_vs = torch.from_numpy(actions[:, indices[k:k+M], ...].reshape(T*M, -1)).float().to(device)
R_vs = torch.from_numpy(Rs[:, indices[k:k+M], ...].reshape(T*M, -1)).float().to(device)
gae_vs = torch.from_numpy(gaes[:, indices[k:k + M], ...].reshape(T*M, -1)).float().to(device)
old_log_prob_action_vs = torch.from_numpy(log_prob_actions[:, indices[k:k+M], ...].reshape(T*M, -1)).float().to(device)
# reconstruct under new policy using old states and old actions
value_vs, logit_vs, _ = net(state_vs, (hx_vs, cx_vs))
dist_vs = Categorical(logits=logit_vs)
log_prob_action_vs = dist_vs.log_prob(action_vs)
log_prob_action_vs = log_prob_action_vs.unsqueeze(1)
adv_v = value_vs - R_vs
ratio_v = torch.exp(log_prob_action_vs - old_log_prob_action_vs)
surr_v = ratio_v * gae_vs
clipped_surr_v = torch.clamp(ratio_v, 1.0 - PPO_EPS, 1.0 + PPO_EPS) * gae_vs
loss_value_v = (0.5 * adv_v.pow(2)).mean()
loss_policy_v = -torch.min(surr_v, clipped_surr_v).mean()
loss_entropy_v = -dist_vs.entropy().mean()
net.zero_grad()
loss_v = COEF_VALUE * loss_value_v + loss_policy_v + COEF_ENTROPY * loss_entropy_v
loss_v.backward()
nn.utils.clip_grad_norm_(net.parameters(), CLIP_GRAD)
optimizer.step()
venv.close()