Distributed TensorFlow Example

Dependencies

• Ubuntu 16.04
• tmux
• TensorFlow >=1.0.0
• Basic knowledge on tmux is required, e.g., shortcut keys.

How to run

• For users with single GPU, do the following command:

$ bash run_single_gpu.sh

In run_single_gpu.sh, you can increase the number of workers by modifying num_workers.

• For users with multiple GPUs, do the following command:

$ bash run_multi_gpu.sh

In run_multi_gpu.sh, you can increase the number of workers by modifying num_workers. Note that the size of GPU_ID in run_multi_gpu.sh should be the same as num_workers. For example, if num_workers is equal to 2, GPU_ID might be (0 1), (2 4), ...

References

• The network architecture from @ischlag. However, not exactly the same. For example, TensorFlow graph is a little bit different.
• Some functions and ideas come from OpenAI's universe-starter-agent. However, the original code does not support GPU usage.