Load from Binary Serializer causes "Object Not Found" error on partial training

[rubixml]

Load from Binary Serializer causes "Object Not Found" error on partial training

[andrewdalpino]

This was a curious issue

It looked as if after the Parameter objects were deserialized from persistence they hashed to a slightly different value even with the Native serializer (off by 1ish) and a more different value with the Binary serializer (off by 7ish)

Rather than having to go through the igbinary guys to see if it was a bug with their extension, I just designed the cache differently and now we don't have to worry about object hashes at all

You can try it out using the current dev-master branch, or you can wait for the next beta release in a few days

Thanks again for the great bug report

[andrewdalpino]

Thank you for the great bug report

This is an interesting one

The bug appears after reconstituting a binary serialized neural network based estimator such as MLP or Logistic Regression with an Adaptive Optimizer (Adam, Momentum, etc.)

Indeed, everything works fine using the Native serializer, and a non-adaptive optimizer (Stochastic, Cyclical, etc.) work ok as well

My gut says that it's something with the igbinary extension (the extension doing the work under the hood)

Specifically, it might have a problem serializing/unserializing the \SplObjectStorage (a hash map structure that uses object hashes as keys)

For example, maybe the object hashes are not being preserved through the deep freeze

That cache you see on line 119 is an \SplObjectStorage that holds the per parameter running velocities and squared gradients needed for the adaptive calculation

\SplObjectStorage is one thing that all the Adaptive Optimizers have in common

I'm going to continue doing some research