Following #54, the mxnet-gluon example will crash (server crashes in RDMA layer) when byteps uses RDMA (DMLC_ENABLE_RDMA=1).

The example initializes byteps (with RDMA), and then creates gluon data loader (code).

# Initialize BytePS
bps.init()

# BytePS: pin context to local rank
context = mx.cpu(bps.local_rank()) if args.no_cuda else mx.gpu(bps.local_rank())
num_workers = bps.size()

# Load training and validation data
train_data, val_data, train_size = get_mnist_iterator()

As explained by @ymjiang in #54, the default (multi-process) gluon data loader forks from the RDMA process which may lead to unexpected errors (reason). Therefore, one very simple solution is to create gluon data loader before bps.init(), which is included in this PR.

Thank you.