def build(self,
input_shape=[None, 784],
custom_variables=None,
scope="optimizee_vars"):
"""Builds the computational graph for the optimziee
:param input_shape: an array with the dimensions of the input
:param custom_variables: an array with variables in the same
order as the networks' variables - this array is used
to build the network with these variables and disconnect
the gradients in the graph.
:returns: cleverhans model for this class
"""
template = MODEL_CONSTANTS.get_fc_template()
mlp_args = {
'layers': template,
'input_shape': input_shape,
'scope': scope
}
if custom_variables is None:
self.model = CustomModel(**mlp_args)
else:
self.model = util.make_with_custom_variables(
CustomModel, mlp_args, custom_variables)
return self.model
def inference_custom():
x = tf.placeholder(dtype=tf.float32, shape=util.get_input_shape_none())
variables = [tf.Variable(initial_value=0)]
inf_op_custom = util.make_with_custom_variables(inference, [x], variables)
return inf_op_custom
def clever_custom_vars():
op = MLP
kwargs = {'layers': layers, 'input_shape': input_shape}
variables, _ = util.get_variables(op, kwargs)
custom_model = util.make_with_custom_variables(op, kwargs, variables)
return custom_model
def get_model_loss(x, y, custom_variables=None):
if custom_variables is None:
model = op(**opkwargs)
logits = model.get_logits(x)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(labels=y,
logits=logits)
return loss
model = util.make_with_custom_variables(op, opkwargs, custom_variables)
logits = model.get_logits(x)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(labels=y, logits=logits)
return loss
def loop_body(t, variables, x, y, fx_array):
with tf.name_scope("loop_function"):
model = util.make_with_custom_variables(op, opkwargs, variables)
logits = model.get_logits(x)
loss = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=y, logits=logits)
fx_array = fx_array.write(t, loss)
with tf.name_scope("time_step"):
t_next = t+1
next_variables = [x+0.01 for x in variables]
return t_next, next_variables, x, y, fx_array