Is your feature request related to a problem? Please describe.
Currently, sending/getting Autograd tensors cause an error because grad_fn is added to the args serialize which is not possible.
Alternative solutions
A. Either we want to serialize AutogradTensor and we A.1 remove the grad_fn arg or A.2 add serialization for all the grad_fn components or B. We don't want AutogradTensor to be sent across the wire and we raise an explicit error
Describe the solution you'd like
We need to serialize grad_fn for AutogradTensor. To do so, one needs to understand how grad_fn is built up in PySyft and how the components should be serialized is an easy way (= try avoiding modifying syft/serde too much)
This triggers weird errors like: https://github.com/OpenMined/PySyft/issues/2666
Related to: https://github.com/OpenMined/PySyft/issues/2137
We should allow autograd tensor to be sendable across the wire, but we should serialize the functions using a string version of them (with specific ones whitelisted) instead of serializaing any function at all.
CC: @karlhigley this is related to what you were talking about.
Can I work on this?
Is this issue still open?
Can I be assigned to work on this issue?
Looks like this got resolved in #2871.
@LaRiffle Please close this issue if it is resolved.
Most helpful comment
We should allow autograd tensor to be sendable across the wire, but we should serialize the functions using a string version of them (with specific ones whitelisted) instead of serializaing any function at all.
CC: @karlhigley this is related to what you were talking about.