I’m not convinced this worked! In you dataset input_fn you didn’t shard the data. That means each of your workers is reading the full dataset!