Conclusion & Demo & Code

In this section, we conclude our project, and list several future research directions. Then we show the demonstration of our system working on the "Softmax computation distribution system". We also give the Github link at the end.

1. Conclusion and Future work.

2. Demo of distributed TensorFlow digit recognition

3. Code on Github

Conclusion and Future work

Conclusion:

In this project, we build Computation distribution & Graph Partition system in TensorFlow environment, and show it is feasible and also can be beneficial if we distribute the compute wisely.

Future work:

— Improve and implement “Auto-Graph Partition” strategy.

— Reduce transmission between different parts of graph. (parameters)

— Use specifically designed model on resource constrained devices

— To improve performance, focus on preprocessing and generality of model

Demo on Youtube

In this demo, the both In-graph replication and Between-graph replication is shown on the cluster. First In-graph replication is shown and later the between graph replication is shown. The input is captured from the camera which is classified using the SoftMax regression model. User decides the node partition and graph allocations across devices.

Code on Github