def policy_model(inp, mdp, spec, name="policy", reuse=None,
track_scope=None):
"""
Predict actions for the given input batch.
Returns:
actions: `batch_size * action_dim`
"""
# TODO remove magic numbers
with tf.variable_scope(name, reuse=reuse,
initializer=tf.truncated_normal_initializer(stddev=0.5)):
return util.mlp(inp, mdp.state_dim, mdp.action_dim,
hidden=spec.policy_dims, track_scope=track_scope)
def critic_model(inp, actions, mdp, spec, name="critic", reuse=None,
track_scope=None):
"""
Predict the Q-value of the given state-action pairs.
Returns:
`batch_size` vector of Q-value predictions.
"""
with tf.variable_scope(name, reuse=reuse):
output = util.mlp(tf.concat(1, [inp, actions]),
mdp.state_dim + mdp.action_dim, 1,
hidden=spec.critic_dims, bias_output=True,
track_scope=track_scope)
return tf.squeeze(output)
def policy_model(inp, mdp, spec, name="policy", reuse=None, track_scope=None):
"""
Predict actions for the given input batch.
Returns:
actions: `batch_size * action_dim`
"""
# TODO remove magic numbers
with tf.variable_scope(
name,
reuse=reuse,
initializer=tf.truncated_normal_initializer(stddev=0.5)):
return util.mlp(inp,
mdp.state_dim,
mdp.action_dim,
hidden=spec.policy_dims,
track_scope=track_scope)
def critic_model(inp,
actions,
mdp,
spec,
name="critic",
reuse=None,
track_scope=None):
"""
Predict the Q-value of the given state-action pairs.
Returns:
`batch_size` vector of Q-value predictions.
"""
with tf.variable_scope(name, reuse=reuse):
output = util.mlp(tf.concat(1, [inp, actions]),
mdp.state_dim + mdp.action_dim,
1,
hidden=spec.critic_dims,
bias_output=True,
track_scope=track_scope)
return tf.squeeze(output)