def get_vars(self):
"""
Provides access to the model's Variables.
This may include Variables that are not parameters, such as batch
norm running moments.
:return: A list of all Variables defining the model.
"""
# Catch eager execution and assert function overload.
try:
if tf.executing_eagerly():
raise NotImplementedError(
"For Eager execution - get_vars " "must be overridden."
)
except AttributeError:
pass
done = False
tried_to_make_params = False
while not done:
# Most models in cleverhans use only trainable variables and do not
# make sure the other collections are updated correctly.
trainable_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, self.scope + "/"
)
# When wrapping other code, such as the CIFAR 10 challenge models,
# we need to make sure we get the batch norm running averages as well
# as the trainable variables.
model_vars = tf.get_collection(
tf.GraphKeys.MODEL_VARIABLES, self.scope + "/"
)
scope_vars = ordered_union(trainable_vars, model_vars)
if len(scope_vars) > 0:
done = True
else:
assert not tried_to_make_params
tried_to_make_params = True
self.make_params()
# Make sure no variables have been added or removed
if hasattr(self, "num_vars"):
assert self.num_vars == len(scope_vars)
else:
self.num_vars = len(scope_vars)
return scope_vars
def get_params(self):
out = []
for layer in self.layers:
out = ordered_union(out, layer.get_params())
return out
