-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
143 lines (116 loc) · 4.84 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
import argparse
import os
import time
from collections import deque
import gym
import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
from agent import PrioritizedDoubleDQNAgent
from model import FCModel
from utils import PrioritizedReplayMemory, beta_adder, cur_time
RPM_SIZE = int(1e6)
RPM_WARMUP_SIZE = RPM_SIZE // 20
UPDATE_FREQ = 4
def run_train_episode(env, agent, rpm, get_beta, n_step, gamma, eps):
obs, done = env.reset(), False
value_losses = [0.0]
step, episode_reward = 0, 0
traj = deque(maxlen=n_step)
while not done:
step += 1
obs = obs.astype('float32')
if np.random.random() < eps:
act = env.action_space.sample()
else:
act = agent.predict(obs)
next_obs, reward, done, info = env.step(act)
traj.append((obs, act, reward, next_obs, done))
if len(traj) == n_step:
g = sum(gamma**i * r for i, (_, _, r, _, _) in enumerate(traj))
rpm.append(obs, act, g, traj[-1][3], traj[-1][4])
if len(rpm) > RPM_WARMUP_SIZE and step % UPDATE_FREQ == 0:
batch, weights, idxes = rpm.sample_batch(32, beta=get_beta())
value_loss, delta = agent.learn(*batch, weights)
value_losses.append(value_loss)
rpm.update_priorities(idxes, delta)
episode_reward += reward
obs = next_obs
g = 0
for i, (obs, act, reward, next_obs, done) in enumerate(reversed(traj)):
g = reward + gamma * g
rpm.append(obs, act, g, traj[-1][3], traj[-1][4])
return episode_reward, np.mean(value_losses)
def run_evaluate_episode(env, agent):
obs, done = env.reset(), False
episode_reward = 0
while not done:
obs = obs.astype('float32')
act = agent.predict(obs)
obs, reward, done, _ = env.step(act)
episode_reward += reward
return episode_reward
def run(args):
torch.manual_seed(args.seed)
log_path = os.path.join(args.out_log, cur_time())
writer = SummaryWriter(log_path)
env = gym.make(args.env)
rpm = PrioritizedReplayMemory(max_size=RPM_SIZE,
obs_size=env.observation_space.shape,
act_size=(1, ),
reward_size=(1, ),
alpha=args.alpha)
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.n
model = FCModel(obs_dim, act_dim)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
agent = PrioritizedDoubleDQNAgent(model, args, device)
get_beta = beta_adder(args.beta)
pbar = tqdm(range(args.num_episodes))
# Annealing epsilon
eps = 0.6
eps_decay_step = (eps - 0.1) / args.num_episodes
for i in pbar:
episode_reward, value_loss = run_train_episode(env,
agent,
rpm,
get_beta,
n_step=args.n_step,
gamma=args.gamma,
eps=eps)
eps = max(0.1, eps - eps_decay_step)
writer.add_scalar('train_episode_reward', episode_reward, i)
writer.add_scalar('value_loss', value_loss, i)
if len(rpm) > RPM_WARMUP_SIZE and (i + 1) % args.eval_freq == 0:
eval_reward = np.mean(
[run_evaluate_episode(env, agent) for _ in range(10)])
writer.add_scalar('evaluate_reward', eval_reward, i)
if (i + 1) % args.sync_period == 0:
agent.sync_target()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--num_episodes',
default=10000,
type=int,
help='number of training episodes')
parser.add_argument('--env',
default='MountainCar-v0',
type=str,
help='name of the gym environment')
parser.add_argument('--out_log',
default='./logs',
type=str,
help='path of tensorboard log file')
parser.add_argument('--seed', default=864, type=int)
# DQN hyperparameters
parser.add_argument('--gamma', default=0.99, type=float)
parser.add_argument('--lr', default=0.00025, type=float)
parser.add_argument('--eval_freq', default=10, type=int)
parser.add_argument('--sync_period', default=5, type=int)
parser.add_argument('--n_step', default=3, type=int)
# PER hyperparameters
parser.add_argument('--alpha', default=0.6, type=float)
parser.add_argument('--beta', default=0.4, type=float)
args = parser.parse_args()
run(args)