Remove the use of generator in dispatcher

[vdutor]

GPflow makes use of a multipledispatch Dispatcher that internally uses a generator. However, according to TensorFlow Capabilities and Limitations generators are not supported by AutoGraph and probably will never be. Thus, compiling code that passed though the dispatcher led to the following warnings:

WARNING:tensorflow:Entity <bound method Dispatcher.dispatch_iter of <dispatched sample_conditional>> appears to be a generator function. It will not be converted by AutoGraph.
WARNING: Entity <bound method Dispatcher.dispatch_iter of <dispatched sample_conditional>> appears to be a generator function. It will not be converted by AutoGraph.
WARNING:tensorflow:Entity <bound method Dispatcher.dispatch_iter of <dispatched conditional>> appears to be a generator function. It will not be converted by AutoGraph.
WARNING: Entity <bound method Dispatcher.dispatch_iter of <dispatched conditional>> appears to be a generator function. It will not be converted by AutoGraph.

This PR still uses the same dispatcher, but overwrites the problematic method that uses python generators and replaces it by a simple list.

NOTE

As of this PR we do not need to write autograph=False in tf.function anymore, and all the code inside dispatching gets compiled the same way as everything else :) !

[vdutor]

@st-- , @awav : addressed all comments. Ready to be merged now.

[awav]

@vdutor +2,068 −47,924. I wish it were actual removing :)

[cdmatters]

2 comments:

1. I worry a bit aboutthe naming of the original package being multipledispatch (class Dispatcher), and our custom file being also called multipledispatch (class Dispatcher). its only a fat finger that changes import .multipledispatch to import multipledispatch! Why doent we change the name of our utils file to gpflow_multipledispatch or custom_multipledispatch
2. Secondly we should have some tests for this diff! A simple test is to create a dispatched function using the two dispatchers, checking the results are the same, but that one generates warnings when wrapped in tf.function!

[awav]

I upvote @condnsdmatters suggestion on testing the warning. Although I don't think we need to test multipledispatch, and we definitely can use with warnings.catch_warnings(record=True) as w: to check that new Dispatch with tf.function gives no warnings.

[vdutor]

Ready for merge once tests pass.

[awav]

Looks good

[st--]

Will merge this once #1265 is in develop

[awav]

@st--, it is a bad practice to work on the branch from another author. Especially, when author gives green light for merging ;) I was about to hit the button.

[st--]

@awav yeah but this branch was off the other one and I wanted to merge them separately