CNN training using TVM compiler

This repo conains simple example how to use TVM compiler for CNN training.

Code realizes a simple LeNet-5 training with SGD solver. The whole training pipline realiseg in single TVM graph:

        batch<-----TrainDB
          |           |
    | -->CNN          |
    |     |           |
    |    loss <--labels
    |     |
    |    grads
    |     |
    |    momentum = momentum factor*momentum + (1.0-momentum_factor)*grads
    |----network_params = network_params - LR*momentum

Training process ends after 1 epoch.

To start process simply run: python lenet_tvm.py

Expected output: TVM:0.44873267 | accuracy:0.90625 | second per iteration:1.5032396518446336