output_nonlinearity=tf.nn.tanh,
)
baseline = NNBaseline(
env_spec=env_spec,
feature_network=cnn,
hidden_sizes=(128, 64),
hidden_nonlinearity=tf.nn.relu,
init_lr=0.001,
n_itr=5,
)
batch_size = 2400
idle = Idle(
env=env,
policy=policy,
baseline=baseline,
)
icm = ICM(
env,
idle,
'/home/fshentu/box3d/trpo_box3d_pixel_v8_tf_idle_normalize_fw_%d' %
seed,
feature_dim=256,
forward_weight=0.2,
inverse_tanh=True,
init_learning_rate=1e-4,
icm_batch_size=128,
replay_pool_size=1000000,
n_updates_per_iter=200,
mdp = TfEnv(normalize(env=GymEnv('Box3dReach-v10',record_video=False, \
log_dir='/tmp/gym_test',record_log=False)))
policy = GaussianMLPPolicy(
"mlp_policy",
env_spec=mdp.spec,
hidden_sizes=(64, 64, 32),
output_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(mdp.spec, )
batch_size = 50000
idle = Idle(
env=mdp,
policy=policy,
baseline=baseline,
n_itr=2000,
)
algorithm = ICM(
mdp,
idle,
"/x/mujoco/tfboard_box3d/trpo_box3d_state_v10_tf_icm_idle_%d" % seed,
feature_dim=mdp.spec.observation_space.flat_dim,
forward_weight=0.1,
external_reward_weight=0.0,
init_learning_rate=1e-4,
n_updates_per_iter=500,
)
run_experiment_lite(algorithm.train(),
mdp = TfEnv(normalize(env=GymEnv('Box3dReach-v2',record_video=False, \
log_dir='/tmp/gym_test',record_log=False)))
policy = GaussianMLPPolicy(
"mlp_policy",
env_spec=mdp.spec,
hidden_sizes=(64, 32),
output_nonlinearity=tf.nn.tanh,
)
baseline = LinearFeatureBaseline(mdp.spec, )
batch_size = 5000
algo = Idle(
env=mdp,
policy=policy,
baseline=baseline,
)
algorithm = ICM(
mdp,
algo,
"/home/dianchen/box3d/trpo_box3d_state_v2_tf_idle_%d" % seed,
feature_dim=mdp.spec.observation_space.flat_dim,
forward_weight=0.2,
external_reward_weight=0.99,
inverse_tanh=True,
init_learning_rate=1e-4,
)
run_experiment_lite(algorithm.train(),
stub(globals())
# Param ranges
seeds = range(5)
for seed in seeds:
mdp = TfEnv(normalize(env=GymEnv('Box3dReach-v8',record_video=False, \
log_dir='/tmp/gym_test',record_log=False)))
policy = UniformControlPolicy(mdp.spec)
baseline = LinearFeatureBaseline(mdp.spec, )
batch_size = 5000
algo = Idle(mdp, policy, baseline, n_itr=1000)
algorithm = ICM(
mdp,
algo,
"/home/dianchen/box3d/trpo_box3d_state_v8_tf_idle_%d" % seed,
feature_dim=mdp.spec.observation_space.flat_dim,
forward_weight=0.2,
external_reward_weight=0.99,
inverse_tanh=True,
init_learning_rate=1e-4,
)
run_experiment_lite(algorithm.train(),
exp_prefix='trpo_box3d_state_v8_tf_idle',
n_parallel=1,
policy = GaussianMLPPolicy(
"mlp_policy",
env_spec=mdp.spec,
hidden_sizes=(64, 64, 32),
output_nonlinearity=tf.nn.tanh,
clip_action=False,
)
baseline = LinearFeatureBaseline(mdp.spec, )
batch_size = 50000
idle = Idle(
env=mdp,
policy=policy,
baseline=baseline,
n_itr=1000,
max_path_length=1000,
batch_size=batch_size,
sampler_cls=BatchSampler,
)
algorithm = ICM(
mdp,
idle,
"/home/fred/tfboard_path/trpo_box3d_state_v17_tf_icm_idle_%d" % seed,
feature_dim=mdp.spec.observation_space.flat_dim,
forward_weight=0.3,
external_reward_weight=0.0,
init_learning_rate=1e-4,
n_updates_per_iter=500,
)