def __init__(
self,
input_shape,
output_shape,
input_queue_capacity=1,
input_queue_name="input",
output_queue_capacity=1,
output_queue_name="output",
):
self.input_shape = input_shape
self.output_shape = output_shape
# input
input_queue = tfe.queue.FIFOQueue(
capacity=input_queue_capacity,
shape=input_shape,
shared_name=input_queue_name)
self.input_placeholder = tfe.define_private_placeholder(shape=input_shape)
self.input_op = input_queue.enqueue(self.input_placeholder)
# output
output_queue = tfe.queue.FIFOQueue(
capacity=output_queue_capacity,
shape=output_shape,
shared_name=output_queue_name)
output = output_queue.dequeue()
self.output0 = output.share0
self.output1 = output.share1
# fixedpoint config
self.modulus = output.backing_dtype.modulus
self.bound = tfe.get_protocol().fixedpoint_config.bound_single_precision
self.scaling_factor = tfe.get_protocol().fixedpoint_config.scaling_factor
def binary_crossentropy(y_true, y_pred):
batch_size = y_true.shape.as_list()[0]
batch_size_inv = 1 / batch_size
out = y_true * get_protocol().log(y_pred)
out += (1 - y_true) * get_protocol().log(1 - y_pred)
out = out.negative()
bce = out.reduce_sum(axis=0) * batch_size_inv
return bce
def _rsqrt(converter, node: Any, inputs: List[str]) -> Any:
x_in = converter.outputs[inputs[0]]
if isinstance(x_in, tf.NodeDef):
tensor = x_in.attr["value"].tensor
shape = [i.size for i in tensor.tensor_shape.dim]
dtype = x_in.attr["dtype"].type
if dtype == tf.float32:
nums = array.array("f", tensor.tensor_content)
elif dtype == tf.float64:
nums = array.array("d", tensor.tensor_content)
else:
raise TypeError("Unsupported dtype for rsqrt")
def inputter_fn():
return tf.constant(1 / np.sqrt(np.array(nums).reshape(shape)))
else:
# XXX this is a little weird but the input into rsqrt is public and
# being used only for batchnorm at the moment
prot = tfe.get_protocol()
# pylint: disable=protected-access
decoded = prot._decode(x_in.value_on_0, True)
# pylint: enable=protected-access
def inputter_fn():
return tf.rsqrt(decoded)
x = tfe.define_public_input(converter.model_provider, inputter_fn)
return x
def test_assign_synchronization(self):
# from https://github.com/tf-encrypted/tf-encrypted/pull/665
tf.reset_default_graph()
tfe.get_protocol().clear_initializers()
prot = tfe.protocol.Pond()
tfe.set_protocol(prot)
def poc(x, y):
x_shares = x.unwrapped
y_shares = y.unwrapped
z_shares = [None, None]
with tf.name_scope("fabricated_test"):
with tf.device(prot.server_0.device_name):
z_shares[0] = x_shares[1] + y_shares[1]
with tf.device(prot.server_1.device_name):
z_shares[1] = x_shares[0] + y_shares[0]
return tfe.protocol.pond.PondPrivateTensor(
prot,
z_shares[0],
z_shares[1],
x.is_scaled)
a = prot.define_private_variable(tf.ones(shape=(1, 1)))
b = prot.define_private_variable(tf.ones(shape=(1, 1)))
op = prot.assign(a, poc(a, b))
with tfe.Session() as sess:
sess.run(tfe.global_variables_initializer())
for _ in range(100):
sess.run(op)
result = sess.run(a.reveal())
assert result == np.array([101.])
def sigmoid(x):
"""Computes sigmoid of x element-wise"""
return get_protocol().sigmoid(x)
def relu(x):
"""Computes relu of x element-wise"""
return get_protocol().relu(x)
def tanh(x):
"""Computes tanh of x element-wise"""
return get_protocol().tanh(x)
def prot(self):
return get_protocol()
def __enter__(self) -> "Protocol":
self.last_protocol = tfe.get_protocol()
tfe.set_protocol(self)
return self
@tfe.local_computation(name_scope='provide_input')
def provide_threshold() -> tf.Tensor:
t = 0.5
a = np.array([t] * 10)
# print (a)
b = tf.constant(a)
return b
@tfe.local_computation('result-receiver', name_scope='receive_output')
def receive_output(average: tf.Tensor) -> tf.Operation:
# simply print average
return tf.print("Result:", average)
inputs = provide_input(player_name='inputter-0')
threshold = provide_threshold(player_name='threshold_inputter')
# result = inputs - threshold
result = inputs
from tf_encrypted import get_protocol
result = get_protocol().relu(result)
result_op = receive_output(result)
# mask =
with tfe.Session() as sess:
sess.run(result_op, tag='prune')