gomoku / DI-engine /dizoo /smac /config /smac_5m6m_collaq_config.py
zjowowen's picture
init space
079c32c
raw
history blame
No virus
2.54 kB
from easydict import EasyDict
agent_num = 5
collector_env_num = 16
evaluator_env_num = 8
main_config = dict(
exp_name='smac_5m6m_collaq_seed0',
env=dict(
map_name='5m_vs_6m',
difficulty=7,
reward_only_positive=True,
mirror_opponent=False,
agent_num=agent_num,
collector_env_num=collector_env_num,
evaluator_env_num=evaluator_env_num,
stop_value=0.999,
n_evaluator_episode=32,
obs_alone=True,
manager=dict(
shared_memory=False,
reset_timeout=6000,
),
),
policy=dict(
model=dict(
agent_num=agent_num,
obs_shape=72,
alone_obs_shape=52,
global_obs_shape=98,
action_shape=12,
hidden_size_list=[128],
attention=True,
self_feature_range=[54, 55], # placeholder 4
ally_feature_range=[34, 54], # placeholder 8*7
attention_size=32,
mixer=True,
lstm_type='gru',
dueling=False,
),
learn=dict(
update_per_collect=20,
batch_size=32,
learning_rate=0.0005,
clip_value=4,
double_q=False,
target_update_theta=0.005,
discount_factor=0.95,
collaq_loss_weight=1.0,
),
collect=dict(
n_episode=32,
unroll_len=10,
env_num=collector_env_num,
),
eval=dict(env_num=evaluator_env_num, evaluator=dict(eval_freq=100, )),
other=dict(
eps=dict(
type='linear',
start=1,
end=0.05,
decay=30000,
),
replay_buffer=dict(
replay_buffer_size=50000,
# (int) The maximum reuse times of each data
max_reuse=1e+9,
max_staleness=1e+9,
),
),
),
)
main_config = EasyDict(main_config)
create_config = dict(
env=dict(
type='smac',
import_names=['dizoo.smac.envs.smac_env'],
),
env_manager=dict(type='subprocess'),
policy=dict(type='collaq'),
collector=dict(type='episode', get_train_sample=True),
)
create_config = EasyDict(create_config)
def train(args):
config = [main_config, create_config]
serial_pipeline(config, seed=args.seed)
if __name__ == '__main__':
from ding.entry import serial_pipeline
serial_pipeline((main_config, create_config), seed=0)