Why TOPI does not support batch_normalization?


i want to make VGG-16 with Batch Normalization(BN) by using TOPI. but TOPI does not support BN but relay has. If I use TOPI, do I have to implement it myself?

Because batch_normalization will be decomposed to a set of add/subtract/multiply ops when lowering a Relay graph to TVM IR.

1 Like

thank you for the reply.

So if i don’t use Relay, i have to implement it myself?

I’m afraid so. Batch norm was removed from TOPI about an year ago: https://github.com/apache/incubator-tvm/pull/2626

1 Like

Okay. I need to implement it myself.

Thank you!

Do we still need Batch norm in training?

Yes, batch norm in training is not supported