We explored the different tensorflow models to understand the cost of computations of graph and different visualization techniques.
We set up multiple workers on same machines and tested the capabilites of distributed tensorflow in one machine.
We have set up tensorflow cluster on different heterogeneous devices and are now testing the graph decomposition in multiple scenario's.
We explored setting up of Tensorflow in Pi devices, and connecting them to the cluster. We successfully set up a running cluster on 5 machines.
We explored the softmax regression model in depth, to understand the graph computations. The MNSIT model also can be executed in realtime on PI cluster, which serves as good option for realtime demo.
We attached PI devices with cameras which helped us to do realtime image processing and softmax model scheduling on the cluster.
We implement our graph partitioning system with AlexNet model, and run it on different cluster settings. Then we analyze the computation and memory requirement of this task, and also proposed the graph partition strategy.
We automated the process of scheduling jobs on cluster via frontend server develped in java and backend server developed in python flask.
We set up heterogeneous cluster in lab, to test the different nature of graph decomposition on the cluster and their different cost of computation.
We concluded the work, by presenting it the Prof. Mani Srivastava, and got the encouraging feedback for future directions.