Add multi_gpu support to AutoModel

[beeb]

Everything is in the title! Would be lovely to be able to add a number of GPUs to use with AutoML / AutoModel.

Thanks!

[mikkokotila]

Makes total sense, will be added :)

[beeb]

When you try to use keras' multi_gpu_model you might find that it gives an Exception TypeError: object of type 'NoneType' has no len() regarding model.outputs. (at least for regression tasks = continuous, did not try other tasks)

Searching for a solution on SO I found this and that.

Hope you can find a solution!

EDIT: it is not a problem for the "dense" networks, only for the other types.

[beeb]

Oh actually I managed to fix it by specifying an input shape for all network types. That's all that was missing. See my second SO link in the reply above.