[TOPI] LRN & L2norm Operator #1051
Conversation
|
cc @sxjscience can you do another round of review? |
|
Need to work on the project in Amazon. I'll give a deep dive into how to implement it better on Thursday. For now, I think we need to support L2Normalization with arbitrary axis. |
|
@sxjscience , shall I follow https://www.tensorflow.org/api_docs/python/tf/nn/l2_normalize for l2norm implementation? |
|
@sxjscience , I have added arbitrary axis support for L2norm. Could you please help to review further. |
|
I think it's good now. We can refactor the implementation later. |
tqchen
left a comment
tqchen
left a comment
There was a problem hiding this comment.
It seems that we can also make lrn axis invariant by providing an axis argument for to specify the channel, can we do that?
|
|
||
| rxk = tvm.reduce_axis((0, size), name='rxk') | ||
| sqr_sum = tvm.compute((b, c, h, w), lambda i, l, j, k: tvm.sum( | ||
| tvm.power(pad_data[i, l + rxk, j, k], 2.0), |
There was a problem hiding this comment.
do not use power to calculate the square, instead use x* x
topi/tests/python/test_topi_lrn.py
Outdated
| sqr_sum[i, j, k, l] = sqr_sum[i, j, k, l] + \ | ||
| (a_np[i, j, k, l + rxl] * \ | ||
| a_np[i, j, k, l + rxl]) | ||
| for i in range(axis0): |
There was a problem hiding this comment.
use broadcasting semantics, instead of the loop
|
Thanks for improving the code during the review process, this is now merged! |
3 participants
LRN operator for caffe alexnet model to use with onnx frontend
@tqchen @sxjscience your review comments from #1000 is addressed here. #1000 got accidentally closed.