def __add__(self, *args, **kwargs):
"""
Here is the version of the add method without the decorator: as you can see
it is much more complicated. However you misght need sometimes to specify
some particular behaviour: so here what to start from :)
"""
if isinstance(args[0], th.Tensor):
data = self.child + args[0].numpy()
obj = PaillierTensor()
obj.child = data
return obj
if isinstance(self.child, th.Tensor):
self.child = self.child.numpy()
# Replace all syft tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
"__add__", self, args, kwargs)
# Send it to the appropriates class and get the response
response = getattr(new_self, "__add__")(*new_args, **new_kwargs)
# Put back SyftTensor on the tensors found in the response
response = hook_args.hook_response("__add__",
response,
wrap_type=type(self))
return response
def method_with_grad(*args, **kwargs):
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
name, self, args, kwargs
)
result = getattr(new_self, name)(*new_args, **new_kwargs)
# Put back SyftTensor on the tensors found in the response
result = hook_args.hook_response(name, result, wrap_type=type(self))
result.grad_fn = grad_fn(self, *args, **kwargs)
return result
def overloaded_native_method(self, *args, **kwargs):
"""
Operate the hooking
"""
if not hasattr(self, "child"): # means that it's not a wrapper
method = getattr(self, f"native_{method_name}")
# Run the native function with the new args
try:
response = method(*args, **kwargs)
except BaseException as e:
# we can make some errors more descriptive with this method
raise route_method_exception(e, self, args, kwargs)
else: # means that there is a wrapper to remove
try:
# Replace all torch tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
method_name, self, args, kwargs
)
except BaseException as e:
# we can make some errors more descriptive with this method
raise route_method_exception(e, self, args, kwargs)
# Send the new command to the appropriate class and get the response
method = getattr(new_self, method_name)
response = method(*new_args, **new_kwargs)
response.parents = (self.id, new_self.id)
# For inplace methods, just directly return self
if syft.framework.is_inplace_method(method_name):
return self
# Put back the wrappers where needed
response = hook_args.hook_response(
method_name,
response,
wrap_type=type(self),
new_self=self,
wrap_args=self.get_class_attributes(),
)
if args:
response.parents = (self, args[0])
else:
response.parents = self
response.command = method_name
return response
def matmul(self, *args, **kwargs):
"""
Hook manually matmul to add the truncation part which is inherent to multiplication
in the fixed precision setting
"""
other = args[0]
if isinstance(other, FixedPrecisionTensor):
assert (
self.precision_fractional == other.precision_fractional
), "In matmul, all args should have the same precision_fractional"
if isinstance(self.child, AdditiveSharingTensor) and isinstance(
other.child, torch.Tensor):
# If we try to matmul a FPT>(wrap)>AST with a FPT>torch.tensor,
# we want to perform AST @ torch.tensor
new_self = self.child
new_args = (other, )
new_kwargs = kwargs
elif isinstance(other.child, AdditiveSharingTensor) and isinstance(
self.child, torch.Tensor):
# If we try to matmul a FPT>torch.tensor with a FPT>(wrap)>AST,
# we swap operators so that we do the same operation as above
new_self = other.child
new_args = (self, )
new_kwargs = kwargs
else:
# Replace all syft tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
"matmul", self, args, kwargs)
# Send it to the appropriate class and get the response
response = getattr(new_self, "matmul")(*new_args, **new_kwargs)
# Put back SyftTensor on the tensors found in the response
response = hook_args.hook_response(
"matmul",
response,
wrap_type=type(self),
wrap_args=self.get_class_attributes())
response %= self.field # Wrap around the field
response = response.truncate(other.precision_fractional)
return response
def _hook_method_args(self, *args, **kwargs):
# Replace all syft tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
attr.__name__, self, args, kwargs)
# Send it to the appropriate class and get the response
response = attr(self, new_self, *new_args, **new_kwargs)
# Put back SyftTensor on the tensors found in the response
response = hook_args.hook_response(
attr.__name__,
response,
wrap_type=type(self),
wrap_args=self.get_class_attributes())
return response
def overloaded_syft_method(self, *args, **kwargs):
"""
Operate the hooking
"""
# Replace all syft tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
attr, self, args, kwargs
)
# Send it to the appropriate class and get the response
response = getattr(new_self, attr)(*new_args, **new_kwargs)
# Put back SyftTensor on the tensors found in the response
response = hook_args.hook_response(
attr, response, wrap_type=type(self), wrap_args=self.get_class_attributes()
)
return response
def manual_add(self, *args, **kwargs):
"""
Here is the version of the add method without the decorator: as you can see
it is much more complicated. However you might need sometimes to specify
some particular behaviour: so here what to start from :)
"""
# Replace all syft tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
"add", self, args, kwargs)
print("Log method manual_add")
# Send it to the appropriate class and get the response
response = getattr(new_self, "add")(*new_args, **new_kwargs)
# Put back SyftTensor on the tensors found in the response
response = hook_args.hook_response("add",
response,
wrap_type=type(self))
return response
def overloaded_attr(self, *args, **kwargs):
"""
Operate the hooking
"""
# Replace all syft tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
attr, self, args, kwargs
)
results = {}
for k, v in new_self.items():
results[k] = v.__getattribute__(attr)(*dispatch(new_args, k), **new_kwargs)
# Put back MultiPointerTensor on the tensors found in the response
response = hook_args.hook_response(
attr, results, wrap_type=MultiPointerTensor, wrap_args=self.get_class_attributes()
)
return response
def mul_and_div(self, other, cmd):
"""
Hook manually mul and div to add the truncation/rescaling part
which is inherent to these operations in the fixed precision setting
"""
changed_sign = False
if isinstance(other, FixedPrecisionTensor):
assert (
self.precision_fractional == other.precision_fractional
), "In mul and div, all args should have the same precision_fractional"
assert self.base == other.base, "In mul and div, all args should have the same base"
if isinstance(other, (int, torch.Tensor, AdditiveSharingTensor)):
new_self = self.child
new_other = other
elif isinstance(other, float):
raise NotImplementedError(
"Can't multiply or divide a FixedPrecisionTensor with a float value"
)
elif isinstance(
self.child,
(AdditiveSharingTensor, MultiPointerTensor)) and isinstance(
other.child, torch.Tensor):
# If operands are FPT>AST and FPT>torch.tensor,
# we want to perform the operation on AST and torch.tensor
if cmd == "mul":
new_self = self.child
elif cmd == "div":
new_self = self.child * self.base**self.precision_fractional
new_other = other
elif isinstance(
other.child,
(AdditiveSharingTensor, MultiPointerTensor)) and isinstance(
self.child, torch.Tensor):
# If operands are FPT>torch.tensor and FPT>AST,
# we swap operators so that we do the same operation as above
if cmd == "mul":
new_self = other.child
new_other = self
elif cmd == "div":
# TODO how to divide by AST?
raise NotImplementedError(
"Division of a FixedPrecisionTensor by an AdditiveSharingTensor not implemented"
)
elif (cmd == "mul"
and isinstance(self.child,
(AdditiveSharingTensor, MultiPointerTensor))
and isinstance(other.child,
(AdditiveSharingTensor, MultiPointerTensor))):
# If we try to multiply a FPT>torch.tensor with a FPT>AST,
# we swap operators so that we do the same operation as above
new_self, new_other, _ = hook_args.unwrap_args_from_method(
"mul", self, other, None)
else:
# Replace all syft tensor with their child attribute
new_self, new_other, _ = hook_args.unwrap_args_from_method(
cmd, self, other, None)
# To avoid problems with negative numbers
# we take absolute value of the operands
# The problems could be 1) bad truncation for multiplication
# 2) overflow when scaling self in division
# sgn_self is 1 when new_self is positive else it's 0
# The comparison is different if new_self is a torch tensor or an AST
sgn_self = (new_self > 0).type(self.torch_dtype)
pos_self = new_self * sgn_self
neg_self = new_self * (sgn_self - 1)
new_self = neg_self + pos_self
# sgn_other is 1 when new_other is positive else it's 0
# The comparison is different if new_other is a torch tensor or an AST
sgn_other = (new_other > 0).type(self.torch_dtype)
pos_other = new_other * sgn_other
neg_other = new_other * (sgn_other - 1)
new_other = neg_other + pos_other
# If both have the same sign, sgn is 1 else it's 0
# To be able to write sgn = 1 - (sgn_self - sgn_other) ** 2,
# we would need to overload the __add__ for operators int and AST.
sgn = -((sgn_self - sgn_other)**2) + 1
changed_sign = True
if cmd == "div":
new_self *= self.base**self.precision_fractional
# Send it to the appropriate class and get the response
response = getattr(new_self, cmd)(new_other)
# Put back SyftTensor on the tensors found in the response
response = hook_args.hook_response(
cmd,
response,
wrap_type=type(self),
wrap_args=self.get_class_attributes())
if not isinstance(other, (int, torch.Tensor, AdditiveSharingTensor)):
if cmd == "mul":
# If operation is mul, we need to truncate
response = response.truncate(self.precision_fractional,
check_sign=False)
if changed_sign:
# Give back its sign to response
pos_res = response * sgn
neg_res = response * (sgn - 1)
response = neg_res + pos_res
return response
def overloaded_native_method(self, *args, **kwargs):
"""
Operate the hooking
"""
if not hasattr(self, "child"): # means that it's not a wrapper
# if self is a natural tensor but the first argument isn't,
# wrap self with the appropriate type and re-run
if len(args) > 0 and hasattr(args[0], "child"):
# if we allow this for PointerTensors it opens the potential
# that we could accidentally serialize and send a tensor in the
# arguments
if not isinstance(args[0].child, PointerTensor):
self = type(args[0].child)().on(self, wrap=True)
args = [args[0]]
return overloaded_native_method(self, *args, **kwargs)
method = getattr(self, f"native_{method_name}")
# Run the native function with the new args
try:
response = method(*args, **kwargs)
except BaseException as e:
# we can make some errors more descriptive with this method
raise route_method_exception(e, self, args, kwargs)
else: # means that there is a wrapper to remove
try:
# Replace all torch tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
method_name, self, args, kwargs
)
except BaseException as e: # if there's a type mismatch, try to fix it!
try:
# if the first argument has no child (meaning it's probably raw data),
# try wrapping it with the type of self. We have to except PointerTensor
# because otherwise it can lead to inadvertently sending data to another
# machine
if not hasattr(args[0], "child") and not isinstance(
self.child, PointerTensor
):
# TODO: add check to make sure this isn't getting around a security class
_args = list()
_args.append(type(self)().on(args[0], wrap=False))
for a in args[1:]:
_args.append(a)
args = _args
# Replace all torch tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
method_name, self, args, kwargs
)
except BaseException as e:
# we can make some errors more descriptive with this method
raise route_method_exception(e, self, args, kwargs)
# Send the new command to the appropriate class and get the response
method = getattr(new_self, method_name)
response = method(*new_args, **new_kwargs)
# For inplace methods, just directly return self
if syft.framework.is_inplace_method(method_name):
return self
# if object is a pointer of pointer, set register to False
if isinstance(self.child, PointerTensor):
wrap_args = {"register": False}
else:
wrap_args = {}
# Put back the wrappers where needed
response = hook_args.hook_response(
method_name, response, wrap_type=type(self), new_self=self, wrap_args=wrap_args
)
return response
def overloaded_native_method(self, *args, **kwargs):
"""
Operate the hooking
"""
if not hasattr(self, "child"): # means that it's not a wrapper
# if self is a natural tensor but the first argument isn't,
# wrap self with the appropriate type and re-run
if len(args) > 0 and hasattr(args[0], "child"):
# if we allow this for PointerTensors it opens the potential
# that we could accidentally serialize and send a tensor in the
# arguments
if not isinstance(args[0].child, PointerTensor):
self = type(args[0].child)().on(self, wrap=True)
args = [args[0]]
return overloaded_native_method(self, *args, **kwargs)
method = getattr(self, f"native_{method_name}")
# Run the native function with the new args
try:
response = method(*args, **kwargs)
except BaseException as e:
# we can make some errors more descriptive with this method
raise route_method_exception(e, self, args, kwargs)
else: # means that there is a wrapper to remove
try:
# Replace all torch tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
method_name, self, args, kwargs
)
except BaseException as e: # if there's a type mismatch, try to fix it!
try:
# if the first argument has no child (meaning it's probably raw data),
# try wrapping it with the type of self. We have to except PointerTensor
# because otherwise it can lead to inadvertently sending data to another
# machine
if not hasattr(args[0], "child") and not isinstance(
self.child, PointerTensor
):
# TODO: add check to make sure this isn't getting around
# a security class
_args = []
_args.append(type(self)().on(args[0], wrap=False))
for a in args[1:]:
_args.append(a)
args = _args
elif isinstance(
self.child, PointerTensor
) and syft.framework.is_inplace_method(method_name):
# under very specific conditions, ie inplace methods containing a
# single argument which is a Tensor, we allow automatic sending of
# this tensor. This is helpful to facilitate utilizing python code
# of other library for remote execution
# so clearly, this allows: pointer += tensor
if isinstance(args[0], FrameworkTensor) and len(args) == 1:
args[0].send_(self.child.location, no_wrap=True)
# Replace all torch tensor with their child attribute
new_self, new_args, new_kwargs = hook_args.unwrap_args_from_method(
method_name, self, args, kwargs
)
except BaseException as e:
# we can make some errors more descriptive with this method
raise route_method_exception(e, self, args, kwargs)
# Send the new command to the appropriate class and get the response
method = getattr(new_self, method_name)
response = method(*new_args, **new_kwargs)
# For inplace methods, just directly return self
if syft.framework.is_inplace_method(method_name):
return self
# Put back the wrappers where needed
response = hook_args.hook_response(
method_name,
response,
wrap_type=type(self),
new_self=self,
wrap_args=self.get_class_attributes(),
)
return response