-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
224 lines (176 loc) · 9.04 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
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
214
215
216
217
218
219
220
221
222
223
224
import os
import time
import dmc2gym
import torch
import utils
from logger import Logger
from replay_buffer import ReplayBuffer
from video import VideoRecorder
import algorithms
from arguments import parse_args
import datetime
torch.backends.cudnn.benchmark = True
def make_env(cfg, test=False):
# per dreamer: https://github.com/danijar/dreamer/blob/02f0210f5991c7710826ca7881f19c64a012290c/wrappers.py#L26
camera_id = 2 if cfg.domain_name == 'quadruped' else 0
if test:
env = dmc2gym.make(domain_name=cfg.domain_name,
task_name=cfg.task_name,
test=test,
seed=cfg.seed,
difficulty=None if cfg.test_difficulty == "None" else cfg.test_difficulty,
dynamic=cfg.test_dynamic,
background_dataset_path=None if cfg.test_background_dataset_path == "None" else os.path.join(os.getcwd(), cfg.test_background_dataset_path),
background_dataset_videos=None if cfg.test_background_dataset_videos == "None" else cfg.test_background_dataset_videos,
background_kwargs=None if cfg.test_background_kwargs == "None" else cfg.test_background_kwargs,
camera_kwargs=None if cfg.test_camera_kwargs == "None" else cfg.test_camera_kwargs,
color_kwargs=None if cfg.test_colour_kwargs == "None" else cfg.test_colour_kwargs,
visualize_reward=False,
from_pixels=True,
height=cfg.image_size,
width=cfg.image_size,
frame_skip=cfg.action_repeat,
camera_id=camera_id,
channels_first=True)
else:
env = dmc2gym.make(domain_name=cfg.domain_name,
task_name=cfg.task_name,
test=test,
seed=cfg.seed,
difficulty=None if cfg.difficulty == "None" else cfg.difficulty,
dynamic=cfg.dynamic,
background_dataset_path=None if cfg.background_dataset_path == "None" else os.path.join(os.getcwd(), cfg.background_dataset_path),
background_dataset_videos=None if cfg.background_dataset_videos == "None" else cfg.background_dataset_videos,
background_kwargs=None if cfg.background_kwargs == "None" else cfg.background_kwargs,
camera_kwargs=None if cfg.camera_kwargs == "None" else cfg.camera_kwargs,
color_kwargs=None if cfg.colour_kwargs == "None" else cfg.colour_kwargs,
visualize_reward=False,
from_pixels=True,
height=cfg.image_size,
width=cfg.image_size,
frame_skip=cfg.action_repeat,
camera_id=camera_id,
channels_first=True)
env = utils.FrameStack(env, k=cfg.frame_stack)
env.seed(cfg.seed)
assert env.action_space.low.min() >= -1
assert env.action_space.high.max() <= 1
return env
class Workspace(object):
def __init__(self, cfg):
algo = f"{cfg.algorithm}_ted_coef{cfg.ted_coef}" if cfg.ted else cfg.algorithm
exp_folder = cfg.exp_name if cfg.exp_name else datetime.date.today().strftime(("%d-%b-%Y"))
self.work_dir = os.path.join(os.getcwd(), cfg.log_dir, exp_folder, algo, str(cfg.seed))
assert not os.path.exists(self.work_dir), 'specified working directory already exists'
os.makedirs(self.work_dir)
print(f'workspace: {self.work_dir}')
self.save_dir = os.path.join(self.work_dir, "trained_models")
utils.write_info(cfg, os.path.join(self.work_dir, 'config.log'))
self.cfg = cfg
self.logger = Logger(self.work_dir,
save_tb=False,
log_frequency=cfg.log_freq,
action_repeat=cfg.action_repeat,
agent=cfg.algorithm)
utils.set_seed_everywhere(cfg.seed)
self.device = torch.device(cfg.device)
self.train_env = make_env(cfg, test=False)
self.env = self.train_env
self.test_env = make_env(cfg, test=True)
action_range = [
float(self.env.action_space.low.min()),
float(self.env.action_space.high.max())
]
self.agent = algorithms.make_agent(self.env.observation_space.shape, self.env.action_space.shape, action_range, cfg)
self.replay_buffer = ReplayBuffer(self.env.observation_space.shape,
self.env.action_space.shape,
cfg.replay_buffer_capacity,
self.cfg.image_pad, self.device,
self.cfg.ted)
self.video_recorder = VideoRecorder(self.work_dir if cfg.save_video else None)
self.step = 0
def evaluate(self):
average_episode_reward = 0
for episode in range(self.cfg.num_eval_episodes):
obs = self.env.reset()
self.video_recorder.init(enabled=(episode == 0))
done = False
episode_reward = 0
episode_step = 0
while not done:
with utils.eval_mode(self.agent):
action = self.agent.act(obs, sample=False)
obs, reward, done, info = self.env.step(action)
self.video_recorder.record(self.env)
episode_reward += reward
episode_step += 1
average_episode_reward += episode_reward
self.video_recorder.save(f'{self.step}.mp4')
average_episode_reward /= self.cfg.num_eval_episodes
self.logger.log('eval/episode_reward', average_episode_reward, self.step)
self.logger.dump(self.step)
def run(self):
episode, episode_reward, episode_step, done = 0, 0, 1, True
start_time = time.time()
total_num_steps = self.cfg.num_train_steps + self.cfg.num_test_steps
while self.step <= total_num_steps:
if done:
if self.step > 0:
self.logger.log('train/duration', time.time() - start_time, self.step)
start_time = time.time()
self.logger.dump(self.step, save=(self.step > self.cfg.num_seed_steps))
# evaluate agent periodically
if self.step % self.cfg.eval_freq == 0:
self.logger.log('eval/episode', episode, self.step)
self.evaluate()
self.logger.log('train/episode_reward', episode_reward, self.step)
obs = self.env.reset()
done = False
episode_reward = 0
episode_step = 0
episode += 1
self.logger.log('train/episode', episode, self.step)
# sample action for data collection
if self.step < self.cfg.num_seed_steps:
action = self.env.action_space.sample()
else:
with utils.eval_mode(self.agent):
action = self.agent.act(obs, sample=True)
# run training update
if self.step >= self.cfg.num_seed_steps:
for _ in range(self.cfg.num_train_iters):
self.agent.update(self.replay_buffer, self.logger, self.step)
if self.step > 0 and self.step % self.cfg.save_freq == 0:
saveables = {
"actor": self.agent.actor.state_dict(),
"critic": self.agent.critic.state_dict(),
"critic_target": self.agent.critic_target.state_dict()
}
if self.cfg.ted:
saveables["ted_classifier"] = self.agent.ted_classifier.state_dict()
save_at = os.path.join(self.save_dir, f"env_step{self.step * self.cfg.action_repeat}")
os.makedirs(save_at, exist_ok=True)
torch.save(saveables, os.path.join(save_at, "models.pt"))
next_obs, reward, done, info = self.env.step(action)
# allow infinite bootstrap
done = float(done)
done_no_max = 0 if episode_step + 1 == self.env._max_episode_steps else done
episode_reward += reward
self.replay_buffer.add(obs, action, reward, next_obs, done, done_no_max, episode)
obs = next_obs
episode_step += 1
self.step += 1
if self.step == self.cfg.num_train_steps:
print("Switching to test env")
self.env = self.test_env
done = True
def main(cfg):
from train import Workspace as W
global workspace
workspace = W(cfg)
start_time = time.time()
workspace.run()
print("total run time: ", time.time()-start_time)
if __name__ == '__main__':
cfg = parse_args()
main(cfg)