def add_test_updates(self, updates, tf_n_mc=None, test_optimizer=adagrad_window,
more_tf_params=None, more_replacements=None,
total_grad_norm_constraint=None):
if more_tf_params is None:
more_tf_params = []
if more_replacements is None:
more_replacements = dict()
tf_target = self(tf_n_mc, more_tf_params=more_tf_params)
tf_target = theano.clone(tf_target, more_replacements, strict=False)
grads = pm.updates.get_or_compute_grads(tf_target, self.obj_params + more_tf_params)
if total_grad_norm_constraint is not None:
grads = pm.total_norm_constraint(grads, total_grad_norm_constraint)
updates.update(
test_optimizer(
grads,
self.test_params +
more_tf_params))
def add_obj_updates(self, updates, obj_n_mc=None, obj_optimizer=adagrad_window,
more_obj_params=None, more_replacements=None,
total_grad_norm_constraint=None):
if more_obj_params is None:
more_obj_params = []
if more_replacements is None:
more_replacements = dict()
obj_target = self(obj_n_mc, more_obj_params=more_obj_params)
obj_target = theano.clone(obj_target, more_replacements, strict=False)
grads = pm.updates.get_or_compute_grads(obj_target, self.obj_params + more_obj_params)
if total_grad_norm_constraint is not None:
grads = pm.total_norm_constraint(grads, total_grad_norm_constraint)
updates.update(
obj_optimizer(
grads,
self.obj_params +
more_obj_params))
if self.op.RETURNS_LOSS:
updates.loss = obj_target