'num_layers': enc_gnn_num_layer,
'model_dim': node_input_dim,
'num_neurons': num_neurons,
'num_relations': num_relations,
'num_head': num_attn_head
})
else:
gnn_conf = MultiStepInputGraphNetworkConfig(
hist_rnn_conf={'input_size': node_input_dim},
hist_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
},
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
})
from sc2rl.utils.reward_funcs import great_victor_with_kill_bonus
from sc2rl.utils.state_process_funcs import process_game_state_to_dgl
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)
reward_name = 'victory_if_zero_enemy'
qnet_conf = MultiStepInputQnetConfig(
multi_step_input_qnet_conf={'exploration_method': 'clustered_random'},
qnet_actor_conf={
'spectral_norm': spectral_norm,
'node_input_dim': node_input_dim
})
if use_attention:
gnn_conf = MultiStepInputNetworkConfig()
else:
gnn_conf = MultiStepInputGraphNetworkConfig(
hist_rnn_conf={'input_size': node_input_dim},
hist_enc_conf={
'spectral_norm': spectral_norm,
'model_dim': node_input_dim
},
curr_enc_conf={
'spectral_norm': spectral_norm,
'model_dim': node_input_dim
})
qnet_conf.gnn_conf = gnn_conf
buffer_conf = NstepInputMemoryConfig(memory_conf={'use_return': True})
brain_conf = QmixBrainConfig(brain_conf={'use_double_q': True},
fit_conf={'tau': 0.9})
sample_spec = buffer_conf.memory_conf['spec']
num_hist_steps = buffer_conf.memory_conf['N']
run_device = 'cpu'
use_attention = False
use_hierarchical_actor = True
num_runners = 1
num_samples = 10
eval_episodes = 20
reward_name = 'victory_if_zero_enemy'
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:
n_hist_feature=n_hist_feature,
n_curr_graph=n_curr_graph,
n_curr_feature=n_curr_feature,
n_maximum_num_enemy=n_maximum_num_enemy,
actions=actions,
rewards=rewards,
dones=dones)
return fit_return_dict
if __name__ == "__main__":
from sc2rl.rl.networks.RelationalGraphNetwork import RelationalGraphNetworkConfig
from sc2rl.rl.networks.FeedForward import FeedForwardConfig
from sc2rl.memory.n_step_memory import NstepInputMemoryConfig
from sc2rl.rl.networks.MultiStepInputGraphNetwork import MultiStepInputGraphNetworkConfig
conf = QmixAgentConf()
actor_conf = ActorModuleConfig()
qnet_conf = MultiStepInputQnetConfig()
qnet_conf.gnn_conf = MultiStepInputGraphNetworkConfig()
mixer_gnn_conf = RelationalGraphNetworkConfig()
mixer_ff_conf = FeedForwardConfig()
brain_conf = QmixActorCriticBrainConfig()
buffer_conf = NstepInputMemoryConfig()
QmixAgent(conf, actor_conf, qnet_conf, mixer_gnn_conf, mixer_ff_conf,
brain_conf, buffer_conf)