def test_train_eval_multiprocess(self):
# test only done for this baselines because the feature is coded in base class in DeepQAgent
tp = TrainingParam()
tp.buffer_size = 100
tp.minibatch_size = 8
tp.update_freq = 32
tp.min_observation = 32
tmp_dir = tempfile.mkdtemp()
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
env_init = grid2op.make("rte_case5_example", test=True)
env = make_multi_env(env_init=env_init, nb_env=2)
li_attr_obs_X = ["prod_p", "load_p", "rho"]
# neural network architecture
observation_size = NNParam.get_obs_size(env, li_attr_obs_X)
sizes = [100, 50, 10] # sizes of each hidden layers
kwargs_archi = {
'observation_size': observation_size,
'sizes': sizes,
'activs':
["relu" for _ in sizes], # all relu activation function
"list_attr_obs": li_attr_obs_X
}
kwargs_converters = {
"all_actions": None,
"set_line_status": False,
"change_bus_vect": True,
"set_topo_vect": False
}
nm_ = "AnneOnymous"
train_dqn(env,
name=nm_,
iterations=100,
save_path=tmp_dir,
load_path=None,
logs_dir=tmp_dir,
training_param=tp,
verbose=False,
kwargs_converters=kwargs_converters,
kwargs_archi=kwargs_archi)
baseline_2 = eval_dqn(env_init,
name=nm_,
load_path=tmp_dir,
logs_path=tmp_dir,
nb_episode=1,
nb_process=1,
max_steps=30,
verbose=False,
save_gif=False)
chronics_class=MultifolderWithCache
)
if env.name == "l2rpn_wcci_2020":
env.chronics_handler.real_data.set_filter(lambda x: re.match(".*Scenario_february_.*$", x) is not None)
env.chronics_handler.real_data.reset()
elif env.name == "l2rpn_case14_sandbox":
# all data can be loaded into memory
# env.chronics_handler.real_data.set_filter(lambda x: True)
env.chronics_handler.real_data.reset()
# env.chronics_handler.real_data.
env_init = env
if args.nb_env > 1:
from l2rpn_baselines.utils import make_multi_env
env = make_multi_env(env_init=env_init, nb_env=int(args.nb_env))
tp = TrainingParam()
# NN training
tp.lr = 1e-5
tp.lr_decay_steps = 300000
tp.minibatch_size = 32 * int(args.nb_env)
tp.update_freq = tp.minibatch_size / 2
# limit the number of time steps played per scenarios
tp.step_increase_nb_iter = None # None to deactivate it
tp.min_iter = None
tp.update_nb_iter = None # once 100 scenarios are solved, increase of "step_increase_nb_iter"
# oversampling hard scenarios
tp.oversampling_rate = None # None to deactivate it