db50bca7fd
* scenarionet training * wandb * train utils * fix callback * run PPO * use pg test * save path * use torch * add dependency * update ignore * update training * large model * use curriculum training * add time to exp name * storage_path * restore * update training * use my key * add log message * check seed * restore callback * restore call bacl * add log message * add logging message * restore ray1.4 * length 500 * ray 100 * wandb * use tf * more levels * add callback * 10 worker * show level * no env horizon * callback result level * more call back * add diffuculty * add mroen stat * mroe stat * show levels * add callback * new * ep len 600 * fix setup * fix stepup * fix to 3.8 * update setup * parallel worker! * new exp * add callback * lateral dist * pg dataset * evaluate * modify config * align config * train single RL * update training script * 100w eval * less eval to reveal * 2000 env eval * new trianing * eval 1000 * update eval * more workers * more worker * 20 worker * dataset to database * split tool! * split dataset * try fix * train 003 * fix mapping * fix test * add waymo tqdm * utils * fix bug * fix bug * waymo * int type * 8 worker read * disable * read file * add log message * check existence * dist 0 * int * check num * suprass warning * add filter API * filter * store map false * new * ablation * filter * fix * update filyter * reanme to from * random select * add overlapping checj * fix * new training sceheme * new reward * add waymo train script * waymo different config * copy raw data * fix bug * add tqdm * update readme * waymo * pg * max lateral dist 3 * pg * crash_done instead of penalty * no crash done * gpu * update eval script * steering range penalty * evaluate * finish pg * update setup * fix bug * test * fix * add on line * train nuplan * generate sensor * udpate training * static obj * multi worker eval * filx bug * use ray for testing * eval! * filter senario * id filter * fox bug * dist = 2 * filter * eval * eval ret * ok * update training pg * test before use * store data=False * collect figures * capture pic --------- Co-authored-by: Quanyi Li <quanyi@bolei-gpu02.cs.ucla.edu>
81 lines
2.5 KiB
Python
81 lines
2.5 KiB
Python
import os.path
|
|
|
|
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
|
|
|
|
if __name__ == '__main__':
|
|
env = ScenarioEnv
|
|
args = get_train_parser().parse_args()
|
|
exp_name = get_exp_name(args)
|
|
stop = int(100_000_000)
|
|
|
|
config = dict(
|
|
env=env,
|
|
env_config=dict(
|
|
# scenario
|
|
start_scenario_index=0,
|
|
num_scenarios=32,
|
|
data_directory=os.path.join(SCENARIONET_DATASET_PATH, "pg"),
|
|
sequential_seed=True,
|
|
|
|
# traffic & light
|
|
reactive_traffic=False,
|
|
no_static_vehicles=True,
|
|
no_light=True,
|
|
static_traffic_object=True,
|
|
|
|
# curriculum training
|
|
curriculum_level=4,
|
|
target_success_rate=0.8,
|
|
|
|
# training
|
|
horizon=None,
|
|
use_lateral_reward=True,
|
|
),
|
|
|
|
# # ===== Evaluation =====
|
|
evaluation_interval=2,
|
|
evaluation_num_episodes=32,
|
|
evaluation_config=dict(env_config=dict(start_scenario_index=32,
|
|
num_scenarios=32,
|
|
sequential_seed=True,
|
|
curriculum_level=1, # turn off
|
|
data_directory=os.path.join(SCENARIONET_DATASET_PATH, "pg"))),
|
|
evaluation_num_workers=2,
|
|
metrics_smoothing_episodes=10,
|
|
|
|
# ===== Training =====
|
|
model=dict(fcnet_hiddens=[512, 256, 128]),
|
|
horizon=600,
|
|
num_sgd_iter=20,
|
|
lr=5e-5,
|
|
rollout_fragment_length=500,
|
|
sgd_minibatch_size=100,
|
|
train_batch_size=4000,
|
|
num_gpus=0.5 if args.num_gpus != 0 else 0,
|
|
num_cpus_per_worker=0.4,
|
|
num_cpus_for_driver=1,
|
|
num_workers=2,
|
|
framework="tf"
|
|
)
|
|
|
|
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=1,
|
|
# 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",
|
|
)
|