[TOPI] topi.nn.batch_norm_inference is broken

[headupinclouds]

The topi.nn.batch_norm_inference python implementation appears to be broken.

  File "/dl/mxnet/3rdparty/tvm/topi/python/topi/nn/batch_norm.py", line 54, in batch_norm_inference
    mean = tvm.compute((C, ), lambda c: moving_mean[c])
NameError: name 'C' is not defined

https://github.com/dmlc/tvm/blob/e4b9f986dab8c48ba109a52106565fc4be6b67c4/topi/python/topi/nn/batch_norm.py#L54-L56

Should the offending lines (mean and var return variables) just be removed here?

C++ prototype:

https://github.com/dmlc/tvm/blob/e4b9f986dab8c48ba109a52106565fc4be6b67c4/topi/include/topi/nn/batch_norm.h#L33-L41

For what it's worth, topi.nn.batch_norm_inference doesn't currently have a unit test, and I don't see any other references to it in the repository:

find /dl/mxnet/3rdparty/tvm/ -name "*.py" | xargs grep batch_norm_inference

/dl/mxnet/3rdparty/tvm/topi/python/topi/nn/batch_norm.py:def batch_norm_inference(data, gamma, beta, moving_mean, moving_var, eps, fix_gamma):

[yzhliu]

I feel batch_norm_inference is no longer used as batch_norm gets decomposed into several scalar operations in simplify_inference pass.

[headupinclouds]

I feel batch_norm_inference is no longer used as batch_norm gets decomposed into several scalar operations in simplify_inference pass

I see. Are you suggesting it should just be removed?

[yzhliu]

Yes, let's do it. would you mind shoot a PR to remove?

[headupinclouds]

Sure, I can put together a PR removing both
tvm/topi/include/topi/nn/batch_norm.h,
tvm/topi/python/topi/nn/batch_norm.py, and any references.

[headupinclouds]

pushed #2626