Fix mgpu init issue with new version of keras

[hxy9243]

Ran into this issue while using keras_experiments as a multi-GPU benchmark with newer version of keras (2.2.0).

RuntimeError: It looks like you are subclassing `Model` and you forgot to call `super(YourClass, self).__init__()`. Always start with this line.

Adding this line seems to solve the issue.

[avolkov1]

Thanks. Let me test this. I recently refactored a few things and updated examples. I did not change the API for ModelMGPU though so it should still work.

I was already calling the super init in line 344: https://github.com/avolkov1/keras_experiments/blob/99cd23118ad731553ef708d9eff480da35db7ec9/keras_exp/multigpu/_multigpu.py#L344

Let me test with latest Keras.

[avolkov1]

Thanks again. I tested this and verified the required fix. I'll merge it in. I added another MGPU class that's a wrapper of Keras multi_gpu_model function that I will push soon.

class ModelKerasMGPU(Model):
    '''
    Wrapper class around "keras.utils.multi_gpu_model". This class enabled
    loading and saving transparently.
    '''
    def __init__(self, ser_model, gpus):  # @IgnorePep8 pylint: disable=super-init-not-called
        pmodel = multi_gpu_model(ser_model, gpus)
        # mimic copy constructor via __dict__ update, hence no super-init
        self.__dict__.update(pmodel.__dict__)
        self._smodel = ser_model

    def __getattribute__(self, attrname):
        '''Override load and save methods to be used from the serial-model. The
        serial-model holds references to the weights in the multi-gpu model.
        '''
        # return Model.__getattribute__(self, attrname)
        if 'load' in attrname or 'save' in attrname:
            return getattr(self._smodel, attrname)

        return super(ModelKerasMGPU, self).__getattribute__(attrname)

That's interesting that with this class wrapper I do not need the super init, probably because of the __dict__. The latest multi_gpu_model seems to perform better than my version. The gpus parameter is an integer or list of integers for GPU ids. You might want to test it. The ModelKerasMGPU class wrapper is to enable model-checkpointing with doesn't work with model returned by the multi_gpu_model function.

[hxy9243]

Good to know. Thank you!