name="baseline_conv_network",
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.flat_dim,
conv_filters=(64, 64, 64, 32),
conv_filter_sizes=((5, 5), (3, 3), (3, 3), (3, 3)),
conv_strides=(3, 3, 3, 2),
conv_pads=('SAME', 'SAME', 'SAME', 'SAME'),
hidden_sizes=(256, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
)
policy = GaussianConvFeaturePolicy(
"conv_feature_policy",
env_spec=env_spec,
feature_network=policy_cnn,
hidden_sizes=(128, 64),
clip_action=False,
)
baseline = NNBaseline(
env_spec=env_spec,
feature_network=baseline_cnn,
hidden_sizes=(128, 64),
hidden_nonlinearity=tf.nn.relu,
init_lr=1e-4,
n_itr=5,
train_feature_network=True,
)
batch_size = 9600
name="conv_feature_network",
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.flat_dim,
conv_filters=(32, 32, 32, 32, 32),
conv_filter_sizes=((3,3),(3,3),(3,3),(3,3), (3,3)),
conv_strides=(2, 2, 2, 2, 2),
conv_pads=('SAME', 'SAME', 'SAME', 'SAME', 'SAME'),
hidden_sizes=(256,),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
)
policy = GaussianConvFeaturePolicy(
"conv_feature_policy",
env_spec=env_spec,
feature_network=cnn,
hidden_sizes=(128,64),
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
algo = TRPO(
