Try a deeper network

When running with a config that matches the https://github.com/google/seq2seq/blob/master/example_configs/nmt_large.yml network I ended up running out of memory on the p2.xlarge GPUs. So I backed off the results a lot.

Should try experimenting with a deeper network to see if that improves performance.

Start with example_configs/nmt_large.yml and the reported p2.xlarge GPU configuration, then reproduce the out-of-memory behavior before experimenting. Done means establishing whether a deeper network improves translation performance without exhausting GPU memory.

Written by the indexing model from the issue text.