Created by: pzelazko-intel

When I use inference transpiler in fluid_benchamark.py, I get much worse test accuracy.

This is result of benchmark with inference transpiler:

python fluid_benchmark.py --device CPU --data_set cifar10 --pass_num 1 --iterations 250  --use_inference_transpiler
...
Pass: 0, Iter: 250, Loss: 1.942829


Total examples: 7840, total time: 153.28633, 51.14611 examples/sed

Pass: 0, Loss: 1.942829 , Test Accuracy: 0.107514

And this is without inference transpiler:

python fluid_benchmark.py --device CPU --data_set cifar10 --pass_num 1 --iterations 250
...
Pass: 0, Iter: 250, Loss: 2.071917


Total examples: 7840, total time: 154.19125, 50.84594 examples/sed

Pass: 0, Loss: 2.071917 , Test Accuracy: 0.256182

This problem is connected with fuse_batch_norm method. fuse_relu_mkldnn works properly.