MXNET_NGRAPH_BACKEND=INTERPRETER does not support training

[Sevin F. Varoglu]

Hit an error "what(): [13:44:54] src/engine/./threaded_engine.h:380: Unsupported op 'BatchNormTrainingRelu'" when running inception-v3 or inception-v4 with MXNET_NGRAPH_BACKEND=INTERPRETER.

MXNET_NGRAPH_BACKEND=INTERPRETER KMP_AFFINITY=granularity=fine,compact,1,0 OMP_NUM_THREADS=56 python example/image-classification/train_imagenet.py --network inception-v3 --batch-size 128 --num-epochs 1 --num-examples 12850 --num-classes 10 --lr 0.4 --lr-factor 0.1 --lr-step-epochs 30,60,90 --initializer normal --image-shape 3,299,299 --data-train=/dataset/imagenet299px/imagenet_train.rec --data-val=/dataset/imagenet299px/imagenet_val.rec

[Matthew Brookhart]

Should be a fairly simple fix, replace this line: https://github.com/NervanaSystems/ngraph-mxnet-bridge/blob/e6f35ebd8f30cc81450ddc0181e20475a728beac/src/ngraph_sgcompiler.cc#L188 with a check on the subgraph backend instead of the mxnet context?

[Ashok Emani]

thanks @mbrookhart , will give it a try.

[Ashok Emani]

tested locally with #515 , seems ok. please reopen if any issues.