Tensorflow 2.16.1 compatibility

[Stikus]

Hello, as you can see in pVACtools issue griffithlab/pVACtools#1084 - mhcflurry and/or https://github.com/KarchinLab/mhcnuggets doesn't work with latest 2.16.1 version of tensorflow.

Exception: An error occurred while calling MHCflurry:
Traceback (most recent call last):
  File "/usr/local/bin/mhcflurry-predict", line 8, in <module>
    sys.exit(run())
  File "/usr/local/lib/python3.10/dist-packages/mhcflurry/predict_command.py", line 204, in run
    predictor = Class1PresentationPredictor.load(models_dir)
  File "/usr/local/lib/python3.10/dist-packages/mhcflurry/class1_presentation_predictor.py", line 956, in load
    affinity_predictor = Class1AffinityPredictor.load(
  File "/usr/local/lib/python3.10/dist-packages/mhcflurry/class1_affinity_predictor.py", line 608, in load
    optimized = result.optimize()
  File "/usr/local/lib/python3.10/dist-packages/mhcflurry/class1_affinity_predictor.py", line 653, in optimize
    Class1NeuralNetwork.merge(
  File "/usr/local/lib/python3.10/dist-packages/mhcflurry/class1_neural_network.py", line 1148, in merge
    configure_tensorflow()
  File "/usr/local/lib/python3.10/dist-packages/mhcflurry/common.py", line 131, in configure_tensorflow
    tensorflow.compat.v1.keras.backend.set_session(session)
AttributeError: module 'keras._tf_keras.keras.backend' has no attribute 'set_session'. Did you mean: 'set_epsilon'?

Looks like some legacy code stopped working - maybe tensorflow version should be pinned until code is fixed?

[timodonnell]

Thanks for letting us know - will fix this

[Stikus]

@timodonnell Thanks for fast answer. I should double-check first - pVACtools still using mhcflurry 2.0.6 - https://github.com/griffithlab/pVACtools/blob/master/setup.py#L90, maybe this problem is already fixed in latest version.

[timodonnell]

Ah ok yes was just seeing that - I think it should be already fixed in the latest.

[timodonnell]

While the set_session issue in your stack trace is already fixed, I'm seeing other failures when running using tensorflow==2.16.1 (stack traces below). So I think we should still release a new mhcflurry with the tensorflow version pinned.

        if self.hyperparameters["learning_rate"] is not None:
            K.set_value(
                self.network().optimizer.lr, self.hyperparameters["learning_rate"]
            )
>       fit_info["learning_rate"] = float(K.get_value(self.network().optimizer.lr))
E       AttributeError: 'RMSprop' object has no attribute 'lr'

mhcflurry/class1_neural_network.py:896: AttributeError

        for i, locally_connected_params in enumerate(locally_connected_layers):
>           current_layer = keras.layers.LocallyConnected1D(
                name="lc_%d" % i, **locally_connected_params
            )(current_layer)
E           AttributeError: module 'tensorflow.keras.layers' has no attribute 'LocallyConnected1D'

mhcflurry/class1_neural_network.py:1365: AttributeError

Also if anyone can suggest the fix for these issues so we can support the latest tf, ideally preserving some compatibility with earlier tf versions, please let me know.

[jday1]

@timodonnell please see my PR for fix with passing tests: #239