def run_experiment(variant):
if variant['env_name'] == 'humanoid-rllab':
env = normalize(HumanoidEnv())
elif variant['env_name'] == 'swimmer-rllab':
env = normalize(SwimmerEnv())
else:
env = normalize(GymEnv(variant['env_name']))
pool = SimpleReplayBuffer(
env_spec=env.spec,
max_replay_buffer_size=variant['max_pool_size'],
)
base_kwargs = dict(
min_pool_size=variant['max_path_length'],
epoch_length=variant['epoch_length'],
n_epochs=variant['n_epochs'],
max_path_length=variant['max_path_length'],
batch_size=variant['batch_size'],
n_train_repeat=variant['n_train_repeat'],
eval_render=False,
eval_n_episodes=1,
)
M = variant['layer_size']
qf = NNQFunction(
env_spec=env.spec,
hidden_layer_sizes=(M, M),
)
df = DFunction(
env_spec=env.spec,
hidden_layer_sizes=[M, M]) # discriminator, input is the actions.
vf = VFunction(env_spec=env.spec, hidden_layer_sizes=[M, M])
policy = StochasticNNPolicy(env_spec=env.spec, hidden_layer_sizes=(M, M))
algorithm = SQL(
base_kwargs=base_kwargs,
env=env,
pool=pool,
qf=qf,
policy=policy,
kernel_fn=adaptive_isotropic_gaussian_kernel,
kernel_n_particles=16,
kernel_update_ratio=0.5,
value_n_particles=16,
td_target_update_interval=1000,
qf_lr=variant['qf_lr'],
policy_lr=variant['policy_lr'],
discount=variant['discount'],
reward_scale=variant['reward_scale'],
save_full_state=False,
df=df,
vf=vf,
df_lr=1e-3,
dist=variant['dist'],
)
algorithm.train()
def run_experiment(variant):
if variant['env_name'] == 'humanoid-rllab':
env = normalize(HumanoidEnv())
elif variant['env_name'] == 'swimmer-rllab':
env = normalize(SwimmerEnv())
elif variant['env_name'] == 'ant-rllab':
env = normalize(AntEnv())
elif variant['env_name'] == 'sawyer-rllab':
env = normalize(SawyerTestEnv())
elif variant['env_name'] == 'arm3Ddisc-rllab':
env = normalize(Arm3dDiscEnv())
else:
env = normalize(GymEnv(variant['env_name']))
pool = SimpleReplayBuffer(
env_spec=env.spec, max_replay_buffer_size=variant['max_pool_size'])
sampler = SimpleSampler(
max_path_length=variant['max_path_length'],
min_pool_size=variant['max_path_length'],
batch_size=variant['batch_size'])
base_kwargs = dict(
epoch_length=variant['epoch_length'],
n_epochs=variant['n_epochs'],
n_train_repeat=variant['n_train_repeat'],
eval_render=False,
eval_n_episodes=1,
sampler=sampler)
M = variant['layer_size']
qf = NNQFunction(env_spec=env.spec, hidden_layer_sizes=(M, M))
policy = StochasticNNPolicy(env_spec=env.spec, hidden_layer_sizes=(M, M))
algorithm = SQL(
base_kwargs=base_kwargs,
env=env,
pool=pool,
qf=qf,
policy=policy,
kernel_fn=adaptive_isotropic_gaussian_kernel,
kernel_n_particles=variant['kernel_particles'],
kernel_update_ratio=variant['kernel_update_ratio'],
value_n_particles=variant['value_n_particles'],
td_target_update_interval=variant['td_target_update_interval'],
qf_lr=variant['qf_lr'],
policy_lr=variant['policy_lr'],
discount=variant['discount'],
reward_scale=variant['reward_scale'],
save_full_state=False)
algorithm.train()