def experiment(variant):
from traffic.make_env import make_env
expl_env = make_env(args.exp_name, **variant['env_kwargs'])
eval_env = make_env(args.exp_name, **variant['env_kwargs'])
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
label_num = expl_env.label_num
label_dim = expl_env.label_dim
if variant['load_kwargs']['load']:
load_dir = variant['load_kwargs']['load_dir']
load_data = torch.load(load_dir + '/params.pkl', map_location='cpu')
policy = load_data['trainer/policy']
vf = load_data['trainer/value_function']
else:
hidden_dim = variant['mlp_kwargs']['hidden']
policy = nn.Sequential(nn.Linear(obs_dim, hidden_dim), nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim), nn.ReLU(),
nn.Linear(hidden_dim, action_dim))
policy = SoftmaxPolicy(policy)
print('parameters: ',
np.sum([p.view(-1).shape[0] for p in policy.parameters()]))
vf = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=1,
)
vf_criterion = nn.MSELoss()
eval_policy = ArgmaxDiscretePolicy(policy, use_preactivation=True)
expl_policy = policy
eval_path_collector = MdpPathCollector(
eval_env,
eval_policy,
)
expl_path_collector = MdpPathCollector(
expl_env,
expl_policy,
)
trainer = PPOTrainer(policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
**variant['trainer_kwargs'])
algorithm = TorchOnlineRLAlgorithm(
trainer=trainer,
exploration_env=expl_env,
evaluation_env=eval_env,
exploration_data_collector=expl_path_collector,
evaluation_data_collector=eval_path_collector,
log_path_function=get_traffic_path_information,
**variant['algorithm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
import sys
from traffic.make_env import make_env
expl_env = make_env(args.exp_name)
eval_env = make_env(args.exp_name)
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
module = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=action_dim,
)
policy = SoftmaxPolicy(module, **variant['policy_kwargs'])
qf1 = Mlp(input_size=obs_dim,
output_size=action_dim,
**variant['qf_kwargs'])
target_qf1 = copy.deepcopy(qf1)
qf2 = Mlp(input_size=obs_dim,
output_size=action_dim,
**variant['qf_kwargs'])
target_qf2 = copy.deepcopy(qf2)
eval_policy = ArgmaxDiscretePolicy(policy, use_preactivation=True)
expl_policy = policy
eval_path_collector = MdpPathCollector(
eval_env,
eval_policy,
)
expl_path_collector = MdpPathCollector(
expl_env,
expl_policy,
)
qf_criterion = nn.MSELoss()
trainer = SACDiscreteTrainer(env=eval_env,
policy=policy,
qf1=qf1,
qf2=qf2,
target_qf1=target_qf1,
target_qf2=target_qf2,
qf_criterion=qf_criterion,
**variant['trainer_kwargs'])
replay_buffer = EnvReplayBuffer(
variant['replay_buffer_size'],
expl_env,
)
algorithm = TorchBatchRLAlgorithm(
trainer=trainer,
exploration_env=expl_env,
evaluation_env=eval_env,
exploration_data_collector=expl_path_collector,
evaluation_data_collector=eval_path_collector,
replay_buffer=replay_buffer,
**variant['algorithm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
import sys
from traffic.make_env import make_env
expl_env = make_env(args.exp_name)
eval_env = make_env(args.exp_name)
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
gb = TrafficGraphBuilder(input_dim=4,
ego_init=torch.tensor([0., 1.]),
other_init=torch.tensor([1., 0.]),
edge_index=torch.tensor([[0, 0, 1, 2],
[1, 2, 0, 0]]))
module = GNNNet(pre_graph_builder=gb,
node_dim=16,
output_dim=action_dim,
post_mlp_kwargs=dict(hidden_sizes=[32]),
num_conv_layers=3)
policy = SoftmaxPolicy(module, **variant['policy_kwargs'])
vf = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=1,
)
vf_criterion = nn.MSELoss()
eval_policy = ArgmaxDiscretePolicy(policy, use_preactivation=True)
expl_policy = policy
eval_path_collector = MdpPathCollector(
eval_env,
eval_policy,
)
expl_path_collector = MdpPathCollector(
expl_env,
expl_policy,
)
trainer = PPOTrainer(policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
**variant['trainer_kwargs'])
algorithm = TorchOnlineRLAlgorithm(
trainer=trainer,
exploration_env=expl_env,
evaluation_env=eval_env,
exploration_data_collector=expl_path_collector,
evaluation_data_collector=eval_path_collector,
**variant['algorithm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
import sys
from traffic.make_env import make_env
expl_env = make_env(args.exp_name,**variant['env_kwargs'])
eval_env = make_env(args.exp_name,**variant['env_kwargs'])
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
from graph_builder_multi import MultiTrafficGraphBuilder
gb = MultiTrafficGraphBuilder(input_dim=4, node_num=expl_env.max_veh_num+1,
ego_init=torch.tensor([0.,1.]),
other_init=torch.tensor([1.,0.]),
)
from gnn_net import GNNNet
gnn = GNNNet(
pre_graph_builder = gb,
node_dim = 16,
num_conv_layers=3)
from layers import SelectLayer
encoders = []
encoders.append(nn.Sequential(gnn,SelectLayer(1,0),nn.ReLU()))
sup_learners = []
for i in range(expl_env.max_veh_num):
sup_learner = nn.Sequential(
gnn,
SelectLayer(1,i+1),
nn.ReLU(),
nn.Linear(16, 2),
)
sup_learner = SoftmaxPolicy(sup_learner, learn_temperature=False)
sup_learners.append(sup_learner)
encoders.append(sup_learner)
decoder = Mlp(input_size=int(16+2*expl_env.max_veh_num),
output_size=action_dim,
hidden_sizes=[],
)
from layers import ConcatLayer
need_gradients = np.array([True]*len(encoders))
if variant['no_gradient']:
need_gradients[1:] = False
policy = nn.Sequential(
ConcatLayer(encoders, need_gradients=list(need_gradients), dim=1),
decoder,
)
policy = SoftmaxPolicy(policy, learn_temperature=False)
vf = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=1,
)
vf_criterion = nn.MSELoss()
eval_policy = ArgmaxDiscretePolicy(policy,use_preactivation=True)
expl_policy = policy
eval_path_collector = MdpPathCollector(
eval_env,
eval_policy,
)
expl_path_collector = MdpPathCollector(
expl_env,
expl_policy,
)
from sup_replay_buffer import SupReplayBuffer
replay_buffer = SupReplayBuffer(
observation_dim = obs_dim,
label_dims = [1]*expl_env.max_veh_num,
max_replay_buffer_size = int(1e6),
)
from rlkit.torch.vpg.ppo_sup import PPOSupTrainer
trainer = PPOSupTrainer(
policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
sup_learners=sup_learners,
replay_buffer=replay_buffer,
**variant['trainer_kwargs']
)
algorithm = TorchOnlineRLAlgorithm(
trainer=trainer,
exploration_env=expl_env,
evaluation_env=eval_env,
exploration_data_collector=expl_path_collector,
evaluation_data_collector=eval_path_collector,
log_path_function = get_traffic_path_information,
**variant['algorithm_kwargs']
)
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
from traffic.make_env import make_env
expl_env = make_env(args.exp_name, **variant['env_kwargs'])
eval_env = make_env(args.exp_name, **variant['env_kwargs'])
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
encoders = []
mlp = Mlp(
input_size=obs_dim,
output_size=32,
hidden_sizes=[
32,
],
)
encoders.append(mlp)
sup_learners = []
for i in range(2):
mlp = Mlp(
input_size=obs_dim,
output_size=2,
hidden_sizes=[
32,
],
)
sup_learner = SoftmaxPolicy(mlp, learn_temperature=False)
sup_learners.append(sup_learner)
encoders.append(sup_learner)
decoder = Mlp(
input_size=int(32 + 2 * 2),
output_size=action_dim,
hidden_sizes=[],
)
module = CombineNet(
encoders=encoders,
decoder=decoder,
no_gradient=variant['no_gradient'],
)
policy = SoftmaxPolicy(module, **variant['policy_kwargs'])
vf = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=1,
)
vf_criterion = nn.MSELoss()
eval_policy = ArgmaxDiscretePolicy(policy, use_preactivation=True)
expl_policy = policy
eval_path_collector = MdpPathCollector(
eval_env,
eval_policy,
)
expl_path_collector = MdpPathCollector(
expl_env,
expl_policy,
)
from sup_replay_buffer import SupReplayBuffer
replay_buffer = SupReplayBuffer(
observation_dim=obs_dim,
label_dims=[1, 1],
max_replay_buffer_size=int(1e6),
)
from rlkit.torch.vpg.ppo_sup import PPOSupTrainer
trainer = PPOSupTrainer(policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
sup_learners=sup_learners,
replay_buffer=replay_buffer,
**variant['trainer_kwargs'])
algorithm = TorchOnlineRLAlgorithm(
trainer=trainer,
exploration_env=expl_env,
evaluation_env=eval_env,
exploration_data_collector=expl_path_collector,
evaluation_data_collector=eval_path_collector,
log_path_function=get_traffic_path_information,
**variant['algorithm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
from traffic.make_env import make_env
expl_env = make_env(args.exp_name,**variant['env_kwargs'])
eval_env = make_env(args.exp_name,**variant['env_kwargs'])
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
label_num = expl_env.label_num
label_dim = expl_env.label_dim
if variant['load_kwargs']['load']:
load_dir = variant['load_kwargs']['load_dir']
load_data = torch.load(load_dir+'/params.pkl',map_location='cpu')
policy = load_data['trainer/policy']
vf = load_data['trainer/value_function']
else:
from graph_builder_multi import MultiTrafficGraphBuilder
gb = MultiTrafficGraphBuilder(input_dim=4, node_num=expl_env.max_veh_num+1,
ego_init=torch.tensor([0.,1.]),
other_init=torch.tensor([1.,0.]),
)
if variant['gnn_kwargs']['attention']:
from gnn_attention_net import GNNAttentionNet
gnn_class = GNNAttentionNet
else:
from gnn_net import GNNNet
gnn_class = GNNNet
gnn = gnn_class(
pre_graph_builder=gb,
node_dim=variant['gnn_kwargs']['node'],
conv_type=variant['gnn_kwargs']['conv_type'],
num_conv_layers=variant['gnn_kwargs']['layer'],
hidden_activation=variant['gnn_kwargs']['activation'],
)
from layers import FlattenLayer, SelectLayer
policy = nn.Sequential(
gnn,
SelectLayer(1,0),
FlattenLayer(),
nn.ReLU(),
nn.Linear(variant['gnn_kwargs']['node'],action_dim)
)
policy = SoftmaxPolicy(policy, learn_temperature=False)
print('parameters: ',np.sum([p.view(-1).shape[0] for p in policy.parameters()]))
vf = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=1,
)
vf_criterion = nn.MSELoss()
eval_policy = ArgmaxDiscretePolicy(policy,use_preactivation=True)
expl_policy = policy
eval_path_collector = MdpPathCollector(
eval_env,
eval_policy,
)
expl_path_collector = MdpPathCollector(
expl_env,
expl_policy,
)
trainer = PPOTrainer(
policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
**variant['trainer_kwargs']
)
algorithm = TorchOnlineRLAlgorithm(
trainer=trainer,
exploration_env=expl_env,
evaluation_env=eval_env,
exploration_data_collector=expl_path_collector,
evaluation_data_collector=eval_path_collector,
log_path_function = get_traffic_path_information,
**variant['algorithm_kwargs']
)
algorithm.to(ptu.device)
algorithm.train()