def experiment(variant):
from simple_sup import SimpleSupEnv
expl_env = SimpleSupEnv(**variant['env_kwars'])
eval_env = SimpleSupEnv(**variant['env_kwars'])
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
hidden_dim = variant['hidden_dim']
encoder = nn.Sequential(
nn.Linear(obs_dim,hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim,hidden_dim),
nn.ReLU(),
)
decoder = nn.Linear(hidden_dim, action_dim)
from layers import ReshapeLayer
sup_learner = nn.Sequential(
nn.Linear(hidden_dim, action_dim),
ReshapeLayer(shape=(1, action_dim)),
)
from sup_softmax_policy import SupSoftmaxPolicy
policy = SupSoftmaxPolicy(encoder, decoder, sup_learner)
print('parameters: ',np.sum([p.view(-1).shape[0] for p in policy.parameters()]))
vf = Mlp(
hidden_sizes=[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 rlkit.torch.vpg.ppo_sup_online import PPOSupOnlineTrainer
trainer = PPOSupOnlineTrainer(
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):
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
encoder = nn.Sequential(
nn.Linear(obs_dim, 32),
nn.ReLU(),
nn.Linear(32, 32),
nn.ReLU(),
)
decoder = nn.Linear(32, action_dim)
from layers import ReshapeLayer
sup_learner = nn.Sequential(
nn.Linear(32, int(label_num * label_dim)),
ReshapeLayer(shape=(label_num, label_dim)),
)
from sup_softmax_policy import SupSoftmaxPolicy
policy = SupSoftmaxPolicy(encoder, decoder, sup_learner)
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 = TRPOTrainer(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):
from simple_sup import SimpleSupEnv
expl_env = SimpleSupEnv(**variant['env_kwars'])
eval_env = SimpleSupEnv(**variant['env_kwars'])
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
encoder = nn.Sequential(
nn.Linear(obs_dim, 16),
nn.ReLU(),
)
decoder = nn.Linear(16, action_dim)
from layers import ReshapeLayer
sup_learner = nn.Sequential(
nn.Linear(16, action_dim),
ReshapeLayer(shape=(1, action_dim)),
)
from sup_softmax_policy import SupSoftmaxPolicy
policy = SupSoftmaxPolicy(encoder, decoder, sup_learner)
vf = Mlp(
hidden_sizes=[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_dim=1,
max_replay_buffer_size=int(1e6),
)
from rlkit.torch.vpg.trpo_sup import TRPOSupTrainer
trainer = TRPOSupTrainer(policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
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,
**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'],
)
encoder = gnn
from layers import FlattenLayer, SelectLayer
decoder = nn.Sequential(
SelectLayer(1, 0), FlattenLayer(), nn.ReLU(),
nn.Linear(variant['gnn_kwargs']['node'], action_dim))
from layers import ReshapeLayer
sup_learner = nn.Sequential(
SelectLayer(1, np.arange(1, expl_env.max_veh_num + 1)),
nn.ReLU(),
nn.Linear(variant['gnn_kwargs']['node'], label_dim),
)
from sup_softmax_policy import SupSoftmaxPolicy
policy = SupSoftmaxPolicy(encoder, decoder, sup_learner)
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,
)
from sup_replay_buffer import SupReplayBuffer
replay_buffer = SupReplayBuffer(
observation_dim=obs_dim,
label_dim=label_num,
max_replay_buffer_size=int(1e6),
)
from rlkit.torch.vpg.ppo_sup_vanilla import PPOSupVanillaTrainer
trainer = PPOSupVanillaTrainer(policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
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:
hidden_dim = variant['mlp_kwargs']['hidden']
encoder = nn.Sequential(
nn.Linear(obs_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(),
)
decoder = nn.Linear(hidden_dim, action_dim)
from layers import ReshapeLayer
sup_learner = nn.Sequential(
nn.Linear(hidden_dim,
int(expl_env.label_num * expl_env.label_dim)),
ReshapeLayer(shape=(expl_env.label_num, expl_env.label_dim)),
)
from sup_softmax_policy import SupSoftmaxPolicy
policy = SupSoftmaxPolicy(encoder, decoder, sup_learner)
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,
)
from sup_replay_buffer import SupReplayBuffer
replay_buffer = SupReplayBuffer(
observation_dim=obs_dim,
label_dim=label_num,
max_replay_buffer_size=int(1e6),
)
from rlkit.torch.vpg.ppo_sup_vanilla import PPOSupVanillaTrainer
trainer = PPOSupVanillaTrainer(policy=policy,
value_function=vf,
vf_criterion=vf_criterion,
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()