This code works, but it is currently very slow. Need to very the op placement on different GPUs to figure out why it is slow.

Ref #44

A few things:

• Variables need to be placed on CPU
• Optimizer ops and preprocessing need to be on CPU

It's probably cleaner to put this into the Estimator class. For example, subclass estimator and add support for model replicas.

See also tensorflow/tensorflow#2126

I get the following error on this branch:

InvalidArgumentError (see above for traceback): Cannot assign a device to node 'save/ShardedFilename_1': Could not satisfy explicit device specification '/device:GPU:1' because no supported kernel for GPU devices is available.
Colocation Debug Info:
Colocation group had the following types and devices:
Identity: CPU
ShardedFilename: CPU
[[Node: save/ShardedFilename_1 = ShardedFilename[_device="/device:GPU:1"](save/StringJoin, save/ShardedFilename_1/shard, save/num_shards)]]

@dennybritz : wanted to know if the branch is usable now. If there are some specific ToDos I can help with the implementation