from sc2rl.rl.agents.MAAC.MultiStepActorCriticAgent import MultiStepActorCriticAgent, MultiStepActorCriticAgentConfig
from sc2rl.rl.brains.MAAC.MultiStepActorCriticBrain import MultiStepActorCriticBrainConfig
from sc2rl.rl.networks.MultiStepInputGraphNetwork import MultiStepInputGraphNetworkConfig
from sc2rl.memory.n_step_memory import NstepInputMemoryConfig
from sc2rl.runners.RunnerManager import RunnerConfig, RunnerManager
if __name__ == "__main__":
map_name = "training_scenario_1"
agent_conf = MultiStepActorCriticAgentConfig()
network_conf = MultiStepInputGraphNetworkConfig()
brain_conf = MultiStepActorCriticBrainConfig()
buffer_conf = NstepInputMemoryConfig()
use_attention = False
num_runners = 5
num_samples = 10
sample_spec = buffer_conf.memory_conf['spec']
num_hist_steps = buffer_conf.memory_conf['N']
agent = MultiStepActorCriticAgent(agent_conf,
network_conf,
brain_conf,
buffer_conf,
use_attention=use_attention)
config = RunnerConfig(map_name=map_name, reward_func=great_victor_with_kill_bonus,
state_proc_func=process_game_state_to_dgl,
'num_neurons': num_neurons,
'num_relations': num_relations
},
curr_enc_conf={
'spectral_norm': spectral_norm,
'num_layers': enc_gnn_num_layer,
'model_dim': node_input_dim,
'use_concat': use_concat_input_gnn,
'num_neurons': num_neurons,
'num_relations': num_relations
})
qnet_conf.gnn_conf = gnn_conf
buffer_conf = NstepInputMemoryConfig(memory_conf={
'use_return': True,
'N': num_hist_time_steps,
'gamma': gamma
})
brain_conf = HierarchicalQmixBrainConfig(brain_conf={
'use_double_q':
use_double_q,
'gamma':
gamma,
'eps':
eps_init,
'eps_gamma':
eps_gamma,
'use_mixer_hidden':
use_mixer_hidden
},
fit_conf={
exp_name = '[S4] scheduler'
qnet_conf = MultiStepInputQnetConfig(
qnet_actor_conf={'spectral_norm': spectral_norm})
actor_conf = MultiStepInputActorConfig()
if use_attention:
gnn_conf = MultiStepInputNetworkConfig()
else:
gnn_conf = MultiStepInputGraphNetworkConfig(
hist_enc_conf={'spectral_norm': spectral_norm},
curr_enc_conf={'spectral_norm': spectral_norm})
qnet_conf.gnn_conf = gnn_conf
actor_conf.gnn_conf = gnn_conf
buffer_conf = NstepInputMemoryConfig(memory_conf={'use_return': True})
brain_conf = QmixActorCriticBrainConfig(brain_conf={'use_double_q': True})
sample_spec = buffer_conf.memory_conf['spec']
num_hist_steps = buffer_conf.memory_conf['N']
run_device = 'cpu'
fit_device = 'cuda' if torch.cuda.is_available() else 'cpu'
if use_attention:
raise NotImplementedError
else:
mixer_gnn_conf = RelationalGraphNetworkConfig(
gnn_conf={'spectral_norm': spectral_norm})
mixer_ff_conf = FeedForwardConfig(
mlp_conf={'spectral_norm': spectral_norm})