Mobilenet prunning is not efficient

[katakissof]

When the mobilenet arch is exported the BN and RELU6 should also change. Are you planning to change this ? Also I noticed that the .pb models after the channel pruning (chn-prn-rmt) have worst inference time (with the benchmark tool) than the originals. Any hint in that ? Finally, are there any plans for supporting mobilenet-SSD v2 in the near future ?