For running single node experiments, run the following:

python driver.py

You can explore the options by passing the --help flag. Use the --use-gpu flag to run the experiments on GPUs. Certain flags have been set by default in the driver.py main() function for ease of experimentation.

For running multi-node experiments, a launcher script has been provided in train.sh. To use it, you must edit the script and set the number of nodes you wish to use (--num-nodes), the address of the master node (--master-addr), the total number of trainer processes desired (--num-trainers), and the network interface TensorPipe should use (setting the env vars GLOO_SOCKET_IFNAME and TP_SOCKET_IFNAME to the output of echo $(ip r | grep default | awk '{print $5}')). Then you can simply run the following on each node:

./train.sh <node_rank>

where <node_rank> is a unique rank for each node (0 ... num_nodes-1).