-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathmain.py
184 lines (150 loc) · 7.75 KB
/
main.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
import argparse
import os
import gym
import numpy as np
import pickle
from tqdm import trange
import visdom
import torch
from policies.generative import Generative
from policies.policy import hard_update
from normalized_actions import NormalizedActions
from noises.ounoise import OrnsteinUhlenbeckActionNoise, NormalActionNoise
from utils import save_model, vis_plot
parser = argparse.ArgumentParser()
parser.add_argument('--env-name', default="HalfCheetah-v2",
help='name of the environment to run')
parser.add_argument('--gamma', type=float, default=0.99, metavar='G',
help='discount factor for reward (default: 0.99)')
parser.add_argument('--tau', type=float, default=0.01, metavar='G',
help='discount factor for model (default: 0.01)')
parser.add_argument('--noise', default='normal', choices=['ou', 'normal'])
parser.add_argument('--noise_scale', type=float, default=0.2, metavar='G', help='(default: 0.2)')
parser.add_argument('--batch_size', type=int, default=64, metavar='N', help='batch size (default: 64)')
parser.add_argument('--num_epochs', type=int, default=None, metavar='N', help='number of epochs (default: None)')
parser.add_argument('--num_epochs_cycles', type=int, default=20, metavar='N')
parser.add_argument('--num_steps', type=int, default=1000000, metavar='N',
help='number of training steps (default: 1000000)')
parser.add_argument('--start_timesteps', type=int, default=10000, metavar='N')
parser.add_argument('--eval_freq', type=int, default=5000, metavar='N')
parser.add_argument('--eval_episodes', type=int, default=100, metavar='N')
parser.add_argument('--train_frequency', type=int, default=2048, metavar='N')
parser.add_argument('--replay_size', type=int, default=50000, metavar='N',
help='size of replay buffer (default: 50000)')
parser.add_argument('--training_actor_samples', type=int, default=16, metavar='N',
help='number of times to sample from the actor for calculating the losses (default: 16)')
parser.add_argument('--visualize', default=False, action='store_true')
parser.add_argument('--experiment_name', default=None, type=str,
help='For multiple different experiments, provide an informative experiment name')
parser.add_argument('--print', default=False, action='store_true')
parser.add_argument('--not_autoregressive', default=False, action='store_true')
parser.add_argument('--q_normalization', type=float, default=0.01,
help='Uniformly smooth the Q function in this range.')
parser.add_argument('--target_policy', type=str, default='exponential', choices=['linear', 'boltzman', 'uniform', 'exponential'],
help='Target policy is constructed based on this operator.')
parser.add_argument('--target_policy_q', type=str, default='min', choices=['min', 'max', 'mean', 'none'],
help='The Q value for each sample is determined based on this operator over the two Q networks.')
parser.add_argument('--temp', type=float, default=1.0, help='Boltzman Temperature for normalizing actions')
args = parser.parse_args()
assert args.training_actor_samples > 0
env = NormalizedActions(gym.make(args.env_name))
eval_env = NormalizedActions(gym.make(args.env_name))
agent = Generative(gamma=args.gamma, tau=args.tau, num_inputs=env.observation_space.shape[0],
action_space=env.action_space, replay_size=args.replay_size, actor_samples=args.training_actor_samples,
q_normalization=args.q_normalization, target_policy=args.target_policy,
target_policy_q=args.target_policy_q, autoregressive=not args.not_autoregressive,
temp=args.temp)
results_dict = {'eval_rewards': [],
'value_losses': [],
'policy_losses': [],
'train_rewards': []
}
base_dir = os.getcwd() + '/models/' + args.env_name + '/'
if args.experiment_name is not None:
base_dir += args.experiment_name + '/'
run_number = 0
while os.path.exists(base_dir + str(run_number)):
run_number += 1
base_dir = base_dir + str(run_number)
os.makedirs(base_dir)
if args.noise == 'ou':
noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(env.action_space.shape[0]),
sigma=float(args.noise_scale) * np.ones(env.action_space.shape[0])
)
elif args.noise == 'normal':
noise = NormalActionNoise(mu=np.zeros(env.action_space.shape[0]),
sigma=float(args.noise_scale) * np.ones(env.action_space.shape[0])
)
else:
noise = None
def reset_noise(a_noise):
if a_noise is not None:
a_noise.reset()
print(base_dir)
state = agent.Tensor([env.reset()])
episode_reward = 0
agent.train()
reset_noise(noise)
if args.visualize:
vis = visdom.Visdom(env=base_dir)
else:
vis = None
episode_timesteps = 0
for step in trange(args.num_steps):
with torch.no_grad():
if step % args.eval_freq == 0:
eval_reward = 0
for test_epoch in range(args.eval_episodes):
done = False
eval_state = agent.Tensor([eval_env.reset()])
while not done:
action = agent.select_action(eval_state)
next_eval_state, reward, done, _ = eval_env.step(action.cpu().numpy()[0])
eval_reward += reward
next_eval_state = agent.Tensor([next_eval_state])
eval_state = next_eval_state
results_dict['eval_rewards'].append((step, eval_reward * 1.0 / args.eval_episodes))
if args.print:
try:
print('env: {0}, run number: {1}, step: {2}, reward: {3}, value loss: {4}, policy loss: {5}'.format(
args.env_name,
run_number,
results_dict['eval_rewards'][-1][0],
results_dict['eval_rewards'][-1][1],
results_dict['value_losses'][-1][1],
results_dict['policy_losses'][-1][1]))
except:
pass
save_model(actor=agent.actor, basedir=base_dir)
with open(base_dir + '/results', 'wb') as f:
pickle.dump(results_dict, f)
if step < args.start_timesteps:
action = torch.Tensor(env.action_space.sample()).to(agent.device).unsqueeze(0)
else:
action = agent.select_action(state, noise)
next_state, reward, done, _ = env.step(action.cpu().numpy()[0])
done_bool = False if episode_timesteps + 1 == env.env._max_episode_steps else done
episode_timesteps += 1
episode_reward += reward
action = agent.Tensor(action)
mask = agent.Tensor([not done_bool])
next_state = agent.Tensor([next_state])
reward = agent.Tensor([reward])
agent.store_transition(state, action, mask, next_state, reward)
state = next_state
if done:
results_dict['train_rewards'].append((step, np.mean(episode_reward)))
episode_reward = 0
episode_timesteps = 0
state = agent.Tensor([env.reset()])
reset_noise(noise)
if len(agent.memory) > args.batch_size and step % args.train_frequency == 0:
value_loss, policy_loss = agent.update_parameters(batch_size=args.batch_size,
number_of_iterations=args.train_frequency)
results_dict['value_losses'].append((step, value_loss))
results_dict['policy_losses'].append((step, policy_loss))
vis_plot(vis, results_dict)
with open(base_dir + '/results', 'wb') as f:
pickle.dump(results_dict, f)
save_model(actor=agent.actor, basedir=base_dir)
env.close()