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})
agent_conf = QmixActorCriticAgentConf(agent_conf={'use_clipped_q': True})
agent = QmixActorCriticAgent(conf=agent_conf,
actor_conf=actor_conf,
qnet_conf=qnet_conf,
mixer_gnn_conf=mixer_gnn_conf,
mixer_ff_conf=mixer_ff_conf,
brain_conf=brain_conf,
buffer_conf=buffer_conf)
agent.to(run_device)
mixer_gnn_conf = RelationalNetworkConfig(
gnn_conf={
'model_dim': mixer_input_dim,
'num_layers': mixer_num_layer,
'num_neurons': num_neurons,
'num_relations': num_relations,
'num_head': num_attn_head,
'use_multi_node_types': False
})
else:
mixer_gnn_conf = RelationalGraphNetworkConfig(
gnn_conf={
'spectral_norm': spectral_norm,
'model_dim': mixer_input_dim,
'num_layers': mixer_num_layer,
'use_concat': use_concat_input_gnn,
'num_neurons': num_neurons,
'use_multi_node_types': False,
'node_update_types': mixer_node_update_types,
'edge_update_types': mixer_edge_update_types,
'use_noisy': use_noisy
})
mixer_ff_conf = FeedForwardConfig(
mlp_conf={
'spectral_norm': spectral_norm,
'input_dimension': mixer_input_dim,
'num_neurons': num_neurons,
'use_noisy': use_noisy
})
sup_mixer_conf = SupQmixerConf(nn_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'
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,
'model_dim': node_input_dim
})
mixer_ff_conf = FeedForwardConfig(mlp_conf={
'spectral_norm': spectral_norm,
'input_dimension': node_input_dim
})
agent_conf = QmixAgentConf(agent_conf={'use_clipped_q': True})
agent = QmixAgent(conf=agent_conf,
qnet_conf=qnet_conf,
mixer_gnn_conf=mixer_gnn_conf,
mixer_ff_conf=mixer_ff_conf,
brain_conf=brain_conf,
buffer_conf=buffer_conf)
fit_device = 'cuda' if torch.cuda.is_available() else 'cpu'
if use_attention:
mixer_gnn_conf = RelationalNetworkConfig(
gnn_conf={
'model_dim': mixer_input_dim,
'num_layers': mixer_num_layer,
'num_neurons': num_neurons,
'num_relations': num_relations,
'num_head': num_attn_head
})
else:
mixer_gnn_conf = RelationalGraphNetworkConfig(
gnn_conf={
'spectral_norm': spectral_norm,
'model_dim': mixer_input_dim,
'num_layers': mixer_num_layer,
'use_concat': use_concat_input_gnn,
'num_neurons': num_neurons
})
mixer_ff_conf = FeedForwardConfig(
mlp_conf={
'spectral_norm': spectral_norm,
'input_dimension': mixer_input_dim,
'num_neurons': num_neurons
})
sup_mixer_conf = SupQmixerConf(nn_conf={
'spectral_norm': spectral_norm,
'input_dimension': mixer_input_dim,
'num_neurons': num_neurons
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)