-
Notifications
You must be signed in to change notification settings - Fork 0
/
rl_train.py
241 lines (213 loc) · 8.12 KB
/
rl_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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
# modified from https://github.com/zhejz/carla-roach/blob/main/train_rl.py
from pathlib import Path
import wandb
import torch as th
from carla_gym.envs import EndlessEnv, LeaderboardEnv
from stable_baselines3.common.vec_env import SubprocVecEnv
from stable_baselines3.common.callbacks import CallbackList
from rl_birdview.utils.rl_birdview_wrapper import RlBirdviewWrapper
from rl_birdview.models.ppo import PPO
from rl_birdview.models.ppo_policy import PpoPolicy
from rl_birdview.models.discriminator import Discriminator
from rl_birdview.utils.wandb_callback import WandbCallback
FAKE_BIRDVIEW = True
RESUME_LAST_TRAIN = True
GAIL = True
RGB_GAIL = False
TRAJ_PLOT = False
reward_configs = {
'hero': {
'entry_point': 'reward.valeo_action:ValeoAction',
'kwargs': {}
}
}
terminal_configs = {
'hero': {
'entry_point': 'terminal.valeo_no_det_px:ValeoNoDetPx',
'kwargs': {}
}
}
env_configs = {
'num_zombie_vehicles': [0, 150],
'num_zombie_walkers': [0, 300],
'weather_group': 'dynamic_1.0'
}
env_eval_configs = {
'weather_group': 'simple',
'routes_group': 'train'
}
multi_env_configs = [
{"host": "192.168.0.6", "port": 2000, 'carla_map': 'Town01'},
{"host": "192.168.0.6", "port": 2002, 'carla_map': 'Town01'},
{"host": "192.168.0.6", "port": 2004, 'carla_map': 'Town01'},
{"host": "192.168.0.6", "port": 2006, 'carla_map': 'Town01'},
{"host": "192.168.0.6", "port": 2008, 'carla_map': 'Town01'},
{"host": "192.168.0.6", "port": 2010, 'carla_map': 'Town01'},
]
multi_env_eval_configs = [
{"host": "192.168.0.6", "port": 2012, 'carla_map': 'Town02'},
]
def get_obs_configs(rgb=False):
obs_configs = {
'hero': {
'speed': {
'module': 'actor_state.speed'
},
'control': {
'module': 'actor_state.control'
},
'velocity': {
'module': 'actor_state.velocity'
},
'birdview': {
'module': 'birdview.chauffeurnet',
'width_in_pixels': 192,
'pixels_ev_to_bottom': 40,
'pixels_per_meter': 5.0,
'history_idx': [-16, -11, -6, -1],
'scale_bbox': True,
'scale_mask_col': 1.0
},
'route_plan': {
'module': 'navigation.waypoint_plan',
'steps': 20
}
}
}
if rgb:
obs_configs['hero'].update({
'gnss': {
'module': 'navigation.gnss'
},
'central_rgb': {
'module': 'camera.rgb',
'fov': 90,
'width': 256,
'height': 144,
'location': [1.2, 0.0, 1.3],
'rotation': [0.0, 0.0, 0.0]
},
'left_rgb': {
'module': 'camera.rgb',
'fov': 90,
'width': 256,
'height': 144,
'location': [1.2, -0.25, 1.3],
'rotation': [0.0, 0.0, -45.0]
},
'right_rgb': {
'module': 'camera.rgb',
'fov': 90,
'width': 256,
'height': 144,
'location': [1.2, 0.25, 1.3],
'rotation': [0.0, 0.0, 45.0]
}
})
return obs_configs
def get_env_wrapper_configs(rgb=True):
env_wrapper_configs = {
'input_states': ['control', 'state', 'vel_xy'],
'acc_as_action': True
}
if rgb:
env_wrapper_configs['input_states'].extend(['linear_speed', 'vec', 'cmd', 'command', 'traj', 'rgb'])
return env_wrapper_configs
def env_maker(env_id, config, obs_configs, env_wrapper_configs, rendering=True):
env = EndlessEnv(obs_configs=obs_configs, reward_configs=reward_configs,
terminal_configs=terminal_configs,
seed=2021, no_rendering=(not rendering), **env_configs, **config)
env = RlBirdviewWrapper(env, **env_wrapper_configs, env_id=env_id, resume=RESUME_LAST_TRAIN)
return env
def env_eval_maker(env_id, config, obs_configs, env_wrapper_configs, rendering=True):
env = LeaderboardEnv(obs_configs=obs_configs, reward_configs=reward_configs,
terminal_configs=terminal_configs,
seed=2021, no_rendering=(not rendering), **env_eval_configs, **config)
env = RlBirdviewWrapper(env, **env_wrapper_configs, env_id=env_id, resume=RESUME_LAST_TRAIN)
return env
if __name__ == '__main__':
generate_rgb = FAKE_BIRDVIEW or RGB_GAIL
obs_configs = get_obs_configs(generate_rgb)
env_wrapper_configs = get_env_wrapper_configs(generate_rgb)
env = SubprocVecEnv([lambda env_id=env_id, config=config: env_maker(env_id, config, obs_configs, env_wrapper_configs, rendering=generate_rgb) for env_id, config in enumerate(multi_env_configs)])
env_eval = SubprocVecEnv([lambda env_id=env_id, config=config: env_eval_maker(env_id, config, obs_configs, env_wrapper_configs, rendering=generate_rgb) for env_id, config in enumerate(multi_env_eval_configs)])
if not RGB_GAIL:
features_extractor_entry_point = 'rl_birdview.models.torch_layers:XtMaCNN'
else:
features_extractor_entry_point = 'rl_birdview.models.torch_layers:RGBXtMaCNN'
policy_kwargs = {
'observation_space': env.observation_space,
'action_space': env.action_space,
'policy_head_arch': [256, 256],
'value_head_arch': [256, 256],
'features_extractor_entry_point': features_extractor_entry_point,
'features_extractor_kwargs': {'states_neurons': [256,256]},
'distribution_entry_point': 'rl_birdview.models.distributions:BetaDistribution',
'fake_birdview': FAKE_BIRDVIEW,
'rgb_gail': RGB_GAIL,
'traj_plot': TRAJ_PLOT
}
discriminator_kwargs = {
'observation_space': env.observation_space,
'action_space': env.action_space,
'batch_size': 256,
'disc_head_arch': [256, 256],
'rgb_gail': RGB_GAIL,
'traj_plot': TRAJ_PLOT
}
output_dir = Path('outputs')
output_dir.mkdir(parents=True, exist_ok=True)
last_checkpoint_path = output_dir / 'checkpoint.txt'
wandb_run_id = None
train_kwargs = {
'initial_learning_rate': 2e-5,
'gail': GAIL,
'n_steps_total': 12288,
'batch_size': 256,
'n_epochs': 20,
'gamma': 0.99,
'gae_lambda': 0.9,
'clip_range': 0.2,
'clip_range_vf': 0.2,
'ent_coef': 0.01,
'explore_coef': 0.05,
'vf_coef': 0.5,
'max_grad_norm': 0.5,
'lr_decay': 0.96,
'use_exponential_lr_decay': True,
'gail_gamma': 0.004,
'gail_gamma_decay': 1.0,
'update_adv': False,
}
if RESUME_LAST_TRAIN:
with open(last_checkpoint_path, 'r') as f:
wb_run_path = f.read()
api = wandb.Api()
wandb_run = api.run(wb_run_path)
wandb_run_id = wandb_run.id
all_ckpts = [ckpt_file for ckpt_file in wandb_run.files() if 'ckpt_latest' in ckpt_file.name]
ckpt_file = all_ckpts[0]
ckpt_file.download(replace=True)
ckpt_file_path = ckpt_file.name
saved_variables = th.load(ckpt_file_path, map_location='cuda')
train_kwargs = saved_variables['train_init_kwargs']
policy = PpoPolicy(**saved_variables['policy_init_kwargs'])
policy.load_state_dict(saved_variables['policy_state_dict'])
discriminator = Discriminator(**saved_variables['discriminator_init_kwargs'])
discriminator.load_state_dict(saved_variables['discriminator_state_dict'])
else:
policy = PpoPolicy(**policy_kwargs)
discriminator = Discriminator(**discriminator_kwargs)
if FAKE_BIRDVIEW:
policy.gan_fake_birdview.pretrain()
agent = PPO(
policy=policy,
discriminator=discriminator,
env=env,
**train_kwargs
)
wb_callback = WandbCallback(env, env_eval, wandb_run_id)
callback = CallbackList([wb_callback])
with open(last_checkpoint_path, 'w') as log_file:
log_file.write(wandb.run.path)
agent.learn(total_timesteps=1e8, callback=callback)