I have run into a problem with setting a minimal loss for a NN to train to. I am running an experiment where I am training several hundred nets of the same layout on the same simple problem. After my first run of it I realized that I had one variable that I should take care off, how good the net was, that is to say how much loss it has. I went to set a minimum loss that the nets would have to hit in my training loop and the error
"SESE FIXME: Loops with multiple exits not handled yet!",
or the error
"SESE FIXME: Imperfect loop exits not handled yet!"
depending on how I try to get the loop to have a none fixed amount of iterations. I read the comments in the code from TFCanoncalizeCFG.cpp(processLoop) Line 239, and I understand that these are simply just not implemented. I forked the branch, followed the TF swift readme, and have been trying to build it, I am getting an error in the build process in swift/lib/SILGen/SILGenBridging.cpp, line 1124.
Assertion failed: (bridgeFnType->getParameters()[0].getConvention() == ParameterConvention::Direct_Owned), function emitBridgedToNativeError
Seeing as how the github README says the current build is passing I must be missing something but I am not sure, anyone have any advice on what the problem is?