LMTCNN-2_1 freeze pb

[yangkenny]

Hi,

First of all, thanks for great work.
I re-train with model_type as LMTCNN-2-1 and freeze model as pb with inference.py.
My target is to use this frozen pb and run with Intel Openvino sdk in x86 machine.
It takes 2 steps to complete. First step is convert pb file into openvino format.
I always got error related to is_training can't be true in batch_norm like below.

FusedBatchNorm doesn't support is_training=True. Node multitaskdpcnn/convdp1/dw_batch_norm/FusedBatchNorm

I also try to modify LMTCNN-2-1 and gives batch_norm_param {is_training} True/False for Training and inference respectively. But still can't make anything different.

However, when i convert all pb files inside androidversion/TF_FACE, convert process goes well without any problem.
Would you help hint how androidversion/TF_FACE/ pb files are frozen?

Thanks a lot