def __init__(self,
env_spec,
hidden_layer_sizes=(100, 100),
num_skills=None):
assert num_skills is not None
Serializable.quick_init(self, locals())
Parameterized.__init__(self)
self._Da = env_spec.action_space.flat_dim
self._Do = env_spec.observation_space.flat_dim
self._obs_pl = tf.placeholder(
tf.float32,
shape=[None, self._Do],
name='observation',
)
self._action_pl = tf.placeholder(
tf.float32,
shape=[None, self._Da],
name='actions',
)
self._name = 'discriminator'
self._input_pls = (self._obs_pl, self._action_pl)
self._layer_sizes = list(hidden_layer_sizes) + [num_skills]
self._output_t = self.get_output_for(*self._input_pls)
def __init__(self, env, dim_g, lr=1E-3, hidden_layers_sizes=(100, 100)):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
assert hasattr(env, 'sample_goal')
self._env = env
self._dim_g = dim_g
self._dim_z = dim_g
self._layer_sizes = list(hidden_layers_sizes) + [2 * self._dim_z]
self._name = 'embedding'
self._lr = lr
self._training_ops = list()
self._create_placeholders()
self._zs, self._log_p_zs, self._mus = \
self.get_output_for(self._goals_pl, reuse=False)
# M x K x Dz, M x K x 1, M x K x Dz
self._fetches = [tf.exp(self._log_p_zs), self._zs, self._mus]
self._create_embedding_update()
# Initialize all uninitialized variables. This prevents initializing
# pre-trained policy and qf and vf variables.
self._sess = tf_utils.get_default_session()
uninit_vars = []
for var in tf.global_variables():
try:
self._sess.run(var)
except tf.errors.FailedPreconditionError:
uninit_vars.append(var)
self._sess.run(tf.variables_initializer(uninit_vars))
def __init__(self, name, input_pls, hidden_layer_sizes):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._name = name
self._input_pls = input_pls
self._layer_sizes = list(hidden_layer_sizes) + [None]
self._output_t = self.get_output_for(*self._input_pls)
def __init__(self, *inputs, name, hidden_layer_sizes):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._name = name
self._inputs = inputs
self._layer_sizes = list(hidden_layer_sizes) + [1]
self._output = self._output_for(*self._inputs)
def __init__(self, inputs, name, hidden_layer_sizes):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._name = name
self._inputs = inputs
self._layer_sizes = list(hidden_layer_sizes) + [1]
self._output = self._output_for(self._inputs)
def __init__(self, env, sess, action_initializer=None):
Parameterized.__init__(self)
if action_initializer is None:
action_initializer = tf.random_uniform_initializer(minval=-0.1,
maxval=0.1)
with tf.variable_scope('action_optimizer'):
self.action = tf.get_variable('planner_action',
[1] + list(env.action_space.shape),
initializer=action_initializer)
self.sess = sess
self.env = env
def __init__(self, name, input_pls, input_pls2, hidden_layer_sizes,
output_nonlinearity=None, with_alpha=True, const_alpha=False):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._name = name
self._input_pls = input_pls
self._input_pls2 = input_pls2
self._layer_sizes = list(hidden_layer_sizes) + [None]
self._output_nonlinearity = output_nonlinearity
self._output_t = self.get_output_for(*self._input_pls)
if with_alpha:
if const_alpha:
self._alpha_t = self.get_const_alpha_for()
else:
self._alpha_t = self.get_alpha_for(*self._input_pls2)
def __init__(self,
name,
input_pls,
hidden_layer_sizes,
output_nonlinearity=None,
batchnorm=False,
dropoutvf_placeholder=None,
dropoutvf_keep_prob=1.0):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._name = name
self._input_pls = input_pls
self._layer_sizes = list(hidden_layer_sizes) + [None]
self._output_nonlinearity = output_nonlinearity
self._batchnorm = batchnorm
self.todropoutvf = (dropoutvf_placeholder is not None)
self.dropoutvf_placeholder = dropoutvf_placeholder
self.dropoutvf_keep_prob = dropoutvf_keep_prob
self._output_t = self.get_output_for(*self._input_pls)
def __init__(self, env_spec, layer_sizes=(128, 16),
output_nonlinearity=None, name='observations_preprocessor'):
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._name = name
self._Do = env_spec.observation_space.flat_dim
obs_ph = tf.placeholder(
tf.float32,
shape=(None, self._Do),
name='observations',
)
self._input_pls = (obs_ph, )
self._layer_sizes = layer_sizes
self._output_nonlinearity = output_nonlinearity
self._output_t = self.get_output_for(obs_ph, reuse=tf.AUTO_REUSE)
def __init__(self, env_spec):
Parameterized.__init__(self)
self._env_spec = env_spec
def __init__(self, output_layers, input_layers=None):
self._output_layers = output_layers
self._input_layers = input_layers
Parameterized.__init__(self)
def __init__(self, env_spec):
Parameterized.__init__(self)
self._env_spec = env_spec