def _create_lagrangian_multipliers(optimizer_dict, doo_ds):
for v in optimizer_dict.state:
print('LM:: {}'.format(v))
lag_mul = []
init_ops = []
for v, der in zip(optimizer_dict.state, doo_ds):
lm = slot_creator.create_slot(v.initialized_value(),
utils.val_or_zero(None, v), 'alpha')
init_op = lm.assign(utils.val_or_zero(der, v))
lag_mul.append(lm)
init_ops.append(init_op)
#from IPython import embed;embed()
#utils.val_or_zero(der, v) assign lag_mul <- utils.val_or_zero ==> op
# merge -> return
#
#lag_mul = [slot_creator.create_slot(v.initialized_value(), utils.val_or_zero(None, v), 'alpha') for v, der
# in zip(optimizer_dict.state, doo_ds)]
#lag_mul = [slot_creator.create_slot(tf.truncated_normal(v.shape, stddev=0.01), utils.val_or_zero(der, v), 'alpha') for v, der
# in zip(optimizer_dict.state, doo_ds)]
[
tf.add_to_collection(utils.GraphKeys.LAGRANGIAN_MULTIPLIERS, lm)
for lm in lag_mul
]
utils.remove_from_collection(utils.GraphKeys.GLOBAL_VARIABLES,
*lag_mul)
# this prevents the 'automatic' initialization with tf.global_variables_initializer.
return lag_mul, init_ops
def _create_lagrangian_multipliers(optimizer_dict, doo_ds):
lag_mul = [slot_creator.create_slot(v.initialized_value(), utils.val_or_zero(der, v), 'alpha') for v, der
in zip(optimizer_dict.state, doo_ds)]
[tf.add_to_collection(utils.GraphKeys.LAGRANGIAN_MULTIPLIERS, lm) for lm in lag_mul]
utils.remove_from_collection(utils.GraphKeys.GLOBAL_VARIABLES, *lag_mul)
# this prevents the 'automatic' initialization with tf.global_variables_initializer.
return lag_mul
def _create_lagrangian_multipliers(optimizer_dict, doo_ds):
lag_mul = [slot_creator.create_slot(v, utils.val_or_zero(der, v), 'alpha') for v, der
in zip(optimizer_dict.state, doo_ds)]
[tf.add_to_collection(utils.GraphKeys.LAGRANGIAN_MULTIPLIERS, lm) for lm in lag_mul]
utils.remove_from_collection(utils.GraphKeys.GLOBAL_VARIABLES, *lag_mul)
# this prevents the 'automatic' initialization with tf.global_variables_initializer.
return lag_mul
def _create_zs(optimizer_dict, hyper, d_init_dynamics_d_hyper):
if d_init_dynamics_d_hyper is None: d_init_dynamics_d_hyper = [None] * len(optimizer_dict)
with tf.variable_scope('Z'):
z = [slot_creator.create_slot(v, utils.val_or_zero(der, v), hyper.op.name) for v, der
in zip(optimizer_dict.state, d_init_dynamics_d_hyper)]
[tf.add_to_collection(utils.GraphKeys.ZS, lm) for lm in z]
# in this case it is completely fine to keep zs into the global variable...
return z
def _create_hypergradient_from_dodh(hyper, doo_dhypers):
"""
Creates one hyper-gradient as a variable. doo_dhypers: initialization, that is the derivative of
the outer objective w.r.t this hyper
"""
hgs = slot_creator.create_slot(hyper, utils.val_or_zero(doo_dhypers, hyper), 'hypergradient')
utils.remove_from_collection(utils.GraphKeys.GLOBAL_VARIABLES, hgs)
return hgs
def _create_z(optimizer_dict, hyper, d_init_dynamics_d_hyper):
if d_init_dynamics_d_hyper is None: d_init_dynamics_d_hyper = [None]*len(optimizer_dict.state)
with tf.variable_scope('Z'):
z = [slot_creator.create_slot(v, utils.val_or_zero(der, v), hyper.op.name) for v, der
in zip(optimizer_dict.state, d_init_dynamics_d_hyper)]
[tf.add_to_collection(utils.GraphKeys.ZS, lm) for lm in z]
# utils.remove_from_collection(utils.GraphKeys.GLOBAL_VARIABLES, *z)
# in this case it is completely fine to keep zs into the global variable...
return z
def _create_hypergradient(hyper, doo_dhypers):
"""
Creates one hyper-gradient as a variable..
:param hyper: the relative hyperparameter
:param doo_dhypers: initialization, that is the derivative of the outer objective w.r.t this hyper
:return:
"""
hgs = slot_creator.create_slot(hyper, utils.val_or_zero(doo_dhypers, hyper), 'hypergradient')
utils.remove_from_collection(utils.GraphKeys.GLOBAL_VARIABLES, hgs)
return hgs
