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):
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=[],
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)
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):
from cartpole import CartPoleEnv
expl_env = CartPoleEnv(mode=2)
eval_env = CartPoleEnv(mode=2)
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.n
# import gym
# expl_env = gym.make('CartPole-v0')
# eval_env = gym.make('CartPole-v0')
# obs_dim = eval_env.observation_space.low.size
# action_dim = eval_env.action_space.n
policy = SoftmaxMlpPolicy(input_size=obs_dim,
output_size=action_dim,
**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)
eval_env = make_env(args.exp_name)
obs_dim = eval_env.observation_space.low.size
action_dim = eval_env.action_space.low.size
policy = TanhGaussianPolicy(
obs_dim=obs_dim,
action_dim=action_dim,
**variant['policy_kwargs'],
)
vf = Mlp(
hidden_sizes=[32, 32],
input_size=obs_dim,
output_size=1,
)
vf_criterion = nn.MSELoss()
eval_policy = MakeDeterministic(policy)
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):
from simple_sup_lstm import SimpleSupLSTMEnv
expl_env = SimpleSupLSTMEnv(**variant['env_kwargs'])
eval_env = SimpleSupLSTMEnv(**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['lstm_kwargs']['hidden_dim']
num_lstm_layers = variant['lstm_kwargs']['num_layers']
node_dim = variant['gnn_kwargs']['node_dim']
node_num = expl_env.node_num
input_node_dim = int(obs_dim / node_num)
a_0 = np.zeros(action_dim)
h1_0 = np.zeros((node_num, hidden_dim * num_lstm_layers))
c1_0 = np.zeros((node_num, hidden_dim * num_lstm_layers))
h2_0 = np.zeros((node_num, hidden_dim * num_lstm_layers))
c2_0 = np.zeros((node_num, hidden_dim * num_lstm_layers))
latent_0 = (h1_0, c1_0, h2_0, c2_0)
from lstm_net import LSTMNet
lstm1_ego = LSTMNet(input_node_dim, action_dim, hidden_dim,
num_lstm_layers)
lstm1_other = LSTMNet(input_node_dim, 0, hidden_dim, num_lstm_layers)
lstm2_ego = LSTMNet(node_dim, 0, hidden_dim, num_lstm_layers)
lstm2_other = LSTMNet(node_dim, 0, hidden_dim, num_lstm_layers)
from graph_builder import TrafficGraphBuilder
gb = TrafficGraphBuilder(
input_dim=hidden_dim,
node_num=node_num,
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=variant['gnn_kwargs']['node_dim'],
conv_type=variant['gnn_kwargs']['conv_type'],
num_conv_layers=variant['gnn_kwargs']['num_layers'],
hidden_activation=variant['gnn_kwargs']['activation'],
)
from gnn_lstm2_net import GNNLSTM2Net
policy_net = GNNLSTM2Net(node_num, gnn, lstm1_ego, lstm1_other,
lstm2_ego, lstm2_other)
from layers import FlattenLayer, SelectLayer
post_net = nn.Sequential(SelectLayer(-2, 0), FlattenLayer(2),
nn.ReLU(), nn.Linear(hidden_dim, action_dim))
from softmax_lstm_policy import SoftmaxLSTMPolicy
policy = SoftmaxLSTMPolicy(
a_0=a_0,
latent_0=latent_0,
obs_dim=obs_dim,
action_dim=action_dim,
lstm_net=policy_net,
post_net=post_net,
)
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,
) # TODO: id is also an input
vf_criterion = nn.MSELoss()
from rlkit.torch.policies.make_deterministic import MakeDeterministic
eval_policy = MakeDeterministic(policy)
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,
recurrent=True,
**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 simple_sup_lstm import SimpleSupLSTMEnv
expl_env = SimpleSupLSTMEnv(**variant['env_kwargs'])
eval_env = SimpleSupLSTMEnv(**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['lstm_kwargs']['hidden_dim']
num_layers = variant['lstm_kwargs']['num_layers']
a_0 = np.zeros(action_dim)
h_0 = np.zeros(hidden_dim * num_layers)
c_0 = np.zeros(hidden_dim * num_layers)
latent_0 = (h_0, c_0)
from lstm_net import LSTMNet
lstm_net = LSTMNet(obs_dim, action_dim, hidden_dim, num_layers)
post_net = torch.nn.Linear(hidden_dim, action_dim)
from softmax_lstm_policy import SoftmaxLSTMPolicy
policy = SoftmaxLSTMPolicy(
a_0=a_0,
latent_0=latent_0,
obs_dim=obs_dim,
action_dim=action_dim,
lstm_net=lstm_net,
post_net=post_net,
)
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()
from rlkit.torch.policies.make_deterministic import MakeDeterministic
eval_policy = MakeDeterministic(policy)
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,
recurrent=True,
**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'],
)
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()