huangfu/scenarionet
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update env (#7)

Browse Source 
* add capture script

* gymnasium API

* training with gymnasium API
This commit is contained in:
Quanyi Li
2023-06-23 19:50:40 +01:00
committed by GitHub
parent 88b4faa00f
commit 5f5a5b9531
14 changed files with 151 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
scenarionet_training/scripts/local_test.py
View File

@@ -1,6 +1,7 @@
import os.path
from metadrive.envs.scenario_env import ScenarioEnv
from metadrive.envs.gym_wrapper import GymEnvWrapper
from scenarionet import SCENARIONET_REPO_PATH, SCENARIONET_DATASET_PATH
from scenarionet_training.train_utils.multi_worker_PPO import MultiWorkerPPO
@@ -13,8 +14,8 @@ if __name__ == '__main__':
stop = int(100_000_000)
config = dict(
env=env,
env_config=dict(
env=GymEnvWrapper,
env_config=dict(inner_class=ScenarioEnv, inner_config=dict(
# scenario
start_scenario_index=0,
num_scenarios=32,
@@ -33,8 +34,7 @@ if __name__ == '__main__':
# training
horizon=None,
use_lateral_reward=True,
),
)),
# # ===== Evaluation =====
evaluation_interval=2,

8
scenarionet_training/scripts/train_nuplan.py
View File

@@ -1,13 +1,13 @@
import os.path
from metadrive.envs.gym_wrapper import GymEnvWrapper
from metadrive.envs.scenario_env import ScenarioEnv
from scenarionet import SCENARIONET_REPO_PATH, SCENARIONET_DATASET_PATH
from scenarionet_training.train_utils.multi_worker_PPO import MultiWorkerPPO
from scenarionet_training.train_utils.utils import train, get_train_parser, get_exp_name
config = dict(
env=ScenarioEnv,
env_config=dict(
env=GymEnvWrapper,
env_config=dict(inner_class=ScenarioEnv, inner_config=dict(
# scenario
start_scenario_index=0,
num_scenarios=40000,
@@ -42,7 +42,7 @@ config = dict(
vehicle_config=dict(side_detector=dict(num_lasers=0))
),
)),
# ===== Evaluation =====
evaluation_interval=15,

8
scenarionet_training/scripts/train_pg.py
View File

@@ -1,14 +1,14 @@
import os.path
from ray.tune import grid_search
from metadrive.envs.scenario_env import ScenarioEnv
from metadrive.envs.gym_wrapper import GymEnvWrapper
from scenarionet import SCENARIONET_REPO_PATH, SCENARIONET_DATASET_PATH
from scenarionet_training.train_utils.multi_worker_PPO import MultiWorkerPPO
from scenarionet_training.train_utils.utils import train, get_train_parser, get_exp_name
config = dict(
env=ScenarioEnv,
env_config=dict(
env=GymEnvWrapper,
env_config=dict(inner_class=ScenarioEnv, inner_config=dict(
# scenario
start_scenario_index=0,
num_scenarios=40000,
@@ -41,7 +41,7 @@ config = dict(
vehicle_config=dict(side_detector=dict(num_lasers=0))
),
)),
# ===== Evaluation =====
evaluation_interval=15,

99
scenarionet_training/scripts/train_waymo.py
View File

@@ -1,13 +1,13 @@
import os.path
from metadrive.envs.gym_wrapper import GymEnvWrapper
from metadrive.envs.scenario_env import ScenarioEnv
from scenarionet import SCENARIONET_REPO_PATH, SCENARIONET_DATASET_PATH
from scenarionet_training.train_utils.multi_worker_PPO import MultiWorkerPPO
from scenarionet_training.train_utils.utils import train, get_train_parser, get_exp_name
config = dict(
env=ScenarioEnv,
env_config=dict(
env=GymEnvWrapper,
env_config=dict(inner_class=ScenarioEnv, inner_config=dict(
# scenario
start_scenario_index=0,
num_scenarios=40000,
@@ -41,55 +41,56 @@ config = dict(
vehicle_config=dict(side_detector=dict(num_lasers=0))
),
)),
# ===== Evaluation =====
evaluation_interval=15,
evaluation_num_episodes=1000,
# TODO (LQY), this is a sample from testset do eval on all scenarios after training!
evaluation_config=dict(env_config=dict(start_scenario_index=0,
num_scenarios=1000,
sequential_seed=True,
curriculum_level=1, # turn off
data_directory=os.path.join(SCENARIONET_DATASET_PATH, "waymo_test"))),
evaluation_num_workers=10,
metrics_smoothing_episodes=10,
# ===== Evaluation =====
evaluation_interval=15,
evaluation_num_episodes=1000,
# TODO (LQY), this is a sample from testset do eval on all scenarios after training!
evaluation_config=dict(env_config=dict(start_scenario_index=0,
num_scenarios=1000,
sequential_seed=True,
curriculum_level=1, # turn off
data_directory=os.path.join(SCENARIONET_DATASET_PATH,
"waymo_test"))),
evaluation_num_workers=10,
metrics_smoothing_episodes=10,
# ===== Training =====
model=dict(fcnet_hiddens=[512, 256, 128]),
horizon=600,
num_sgd_iter=20,
lr=1e-4,
rollout_fragment_length=500,
sgd_minibatch_size=200,
train_batch_size=50000,
num_gpus=0.5,
num_cpus_per_worker=0.3,
num_cpus_for_driver=1,
num_workers=20,
framework="tf"
)
# ===== Training =====
model=dict(fcnet_hiddens=[512, 256, 128]),
horizon=600,
num_sgd_iter=20,
lr=1e-4,
rollout_fragment_length=500,
sgd_minibatch_size=200,
train_batch_size=50000,
num_gpus=0.5,
num_cpus_per_worker=0.3,
num_cpus_for_driver=1,
num_workers=20,
framework="tf"
)
if __name__ == '__main__':
# PG data is generated with seeds 10,000 to 60,000
# PG data is generated with seeds 10,000 to 60,000
args = get_train_parser().parse_args()
exp_name = get_exp_name(args)
stop = int(100_000_000)
config["num_gpus"] = 0.5 if args.num_gpus != 0 else 0
exp_name = get_exp_name(args)
stop = int(100_000_000)
config["num_gpus"] = 0.5 if args.num_gpus != 0 else 0
train(
MultiWorkerPPO,
exp_name=exp_name,
save_dir=os.path.join(SCENARIONET_REPO_PATH, "experiment"),
keep_checkpoints_num=5,
stop=stop,
config=config,
num_gpus=args.num_gpus,
# num_seeds=args.num_seeds,
num_seeds=5,
test_mode=args.test,
# local_mode=True,
# TODO remove this when we release our code
# wandb_key_file="~/wandb_api_key_file.txt",
wandb_project="scenarionet",
)
train(
MultiWorkerPPO,
exp_name=exp_name,
save_dir=os.path.join(SCENARIONET_REPO_PATH, "experiment"),
keep_checkpoints_num=5,
stop=stop,
config=config,
num_gpus=args.num_gpus,
# num_seeds=args.num_seeds,
num_seeds=5,
test_mode=args.test,
# local_mode=True,
# TODO remove this when we release our code
# wandb_key_file="~/wandb_api_key_file.txt",
wandb_project="scenarionet",
)