def experiment(variant):
env_params = variant['env_params']
env = SawyerXYZReachingEnv(**env_params)
es = OUStrategy(action_space=env.action_space)
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
qf = FlattenMlp(
input_size=obs_dim + action_dim,
output_size=1,
hidden_sizes=[400, 300],
)
policy = TanhMlpPolicy(
input_size=obs_dim,
output_size=action_dim,
hidden_sizes=[400, 300],
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = DDPG(env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
def example(variant):
env_class = variant['env_class']
env_params = variant['env_params']
env = env_class(**env_params)
normalize(env)
es_class = variant['es_class']
es_params = dict(action_space=env.action_space, **variant['es_params'])
use_gpu = variant['use_gpu']
es = es_class(**es_params)
qf = FeedForwardQFunction(
int(env.observation_space.flat_dim),
int(env.action_space.flat_dim),
100,
100,
)
policy_class = variant['policy_class']
policy_params = dict(
obs_dim=get_dim(env.observation_space),
action_dim=get_dim(env.action_space),
fc1_size=100,
fc2_size=100,
)
policy = policy_class(**policy_params)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
algorithm = DDPG(
env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
**variant['algo_params'],
)
if use_gpu and ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
def experiment(variant):
env_params = variant['env_params']
es_params = variant['es_params']
env = SawyerXYZReachingEnv(**env_params)
es = OUStrategy(action_space=env.action_space, **es_params)
hidden_sizes = variant['hidden_sizes']
obs_dim = env.observation_space.low.size
action_dim = env.action_space.low.size
qf = FlattenMlp(
input_size=obs_dim + action_dim,
output_size=1,
hidden_sizes=[hidden_sizes, hidden_sizes],
)
policy = TanhMlpPolicy(
input_size=obs_dim,
output_size=action_dim,
hidden_sizes=[hidden_sizes, hidden_sizes],
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
if variant['env_params']['relative_pos_control']:
variant['algo_params']['max_path_length'] = 3
variant['algo_params']['num_steps_per_epoch'] = 15
variant['algo_params']['num_steps_per_eval'] = 15
algorithm = DDPG(env,
qf=qf,
policy=policy,
exploration_policy=exploration_policy,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()