I misphrased that point, agreed. It's not classic driver-driven synchronous training, as you would do in tensorflow. It's using all-reduce (not available in ten

I misphrased that point, agreed. It's not classic driver-driven synchronous training, as you would do in tensorflow. It's using all-reduce (not available in tensorflow yet, i think).