def __init__(self,
dim_input,
dim_output,
dim_hidden=32,
num_layers=4,
num_particles=2,
max_test_step=5):
# model size
self.dim_input = dim_input
self.dim_output = dim_output
self.dim_hidden = dim_hidden
self.num_layers = num_layers
self.num_particles = num_particles
# learning rate
self.follow_lr = tf.placeholder_with_default(input=FLAGS.follow_lr,
name='follow_lr',
shape=[])
self.leader_lr = tf.placeholder_with_default(input=FLAGS.leader_lr,
name='leader_lr',
shape=[])
self.meta_lr = tf.placeholder_with_default(input=FLAGS.meta_lr,
name='meta_lr',
shape=[])
# for test time
self.max_test_step = max_test_step
# build model
self.bnn = BNN(dim_input=self.dim_input,
dim_output=self.dim_output,
dim_hidden=self.dim_hidden,
num_layers=self.num_layers,
is_bnn=True)
# init model
self.construct_network_weights = self.bnn.construct_network_weights
# forwarding
self.forward_network = self.bnn.forward_network
# init input data
self.follow_x = tf.placeholder(dtype=tf.float32, name='follow_x')
self.follow_y = tf.placeholder(dtype=tf.float32, name='follow_y')
self.leader_x = tf.placeholder(dtype=tf.float32, name='leader_x')
self.leader_y = tf.placeholder(dtype=tf.float32, name='leader_y')
self.valid_x = tf.placeholder(dtype=tf.float32, name='valid_x')
self.valid_y = tf.placeholder(dtype=tf.float32, name='valid_y')
# init parameters
self.W_network_particles = None
def run_cartpole_expl():
env = gym.make('CartPole-v0')
obs_dim = np.prod(env.observation_space.shape)
act_dim = np.prod(env.action_space.shape)
n_actions = env.action_space.n
policy_hidden_dim = 256
policy = Policy(obs_dim, policy_hidden_dim, n_actions)
input_dim = int(obs_dim + act_dim)
output_dim = int(obs_dim)
hidden_dim = 64
model = BNN(input_dim, hidden_dim, output_dim)
exp = Experiment(policy,
model,
env,
exp_name="cartpole_expl",
train_model=True,
calc_inf_gain=True)
exp.train()
def init_estimator(self, W=None, B=None):
return BNN(layers=self.layers, W=W, B=B)
