We need to support Clip operator from Caffe2. Spec can be found here:
https://caffe2.ai/docs/operators-catalogue.html#clip
artemrakhov-glow
All 9 comments
@yzhao30 feel free to grab this one, let us know if you have any questions.
rdzhabarov
on 25 Sep 2018
Some guidelines:
The code which parses Caffe2 proto leaves here:
https://github.com/pytorch/glow/blob/master/lib/Importer/Caffe2.cpp
When seeing an operator, this code usually calls corresponding method of the Graph (e.g. G_.createClip()). Each Graph method either creates a new node of required type. Or does "immediate lowering": creates node of another type, performing required action. See createFlatten() for example. The same needs to happen here for createClip. Clip could be represented as a combination of Compares, Selects and Splats.
Related issue is supporting ReplaceNaN (#1704 ). It is very similar operator, maybe even a bit simpler.
artemrakhov-glow
on 25 Sep 2018
As a followup would be great to have a documentation on how to implement ONNX/C2 operator in Glow.
Specifically, what are the necessary steps and considerations to have to implement an Op.
rdzhabarov
on 25 Sep 2018
@yzhao30 For testing, you could take a look at this file :
https://github.com/pytorch/glow/blob/master/tests/unittests/caffe2ImporterTest.cpp
beicy
on 26 Sep 2018
Great work!
rdzhabarov
on 28 Sep 2018
As a followup would be great to have a documentation on how to implement ONNX/C2 operator in Glow.
Specifically, what are the necessary steps and considerations to have to implement an Op.
@yzhao30 would you like to tackle this as well?
rdzhabarov
on 28 Sep 2018
As a followup would be great to have a documentation on how to implement ONNX/C2 operator in Glow.
Specifically, what are the necessary steps and considerations to have to implement an Op.@yzhao30 would you like to tackle this as well?
Sure. I can. But this Clip op is just a combination of existing ones. Should we also document on other cases that involve lower layers?
yzhao30
on 29 Sep 2018
@yzhao30 Yes, would be great to cover both cases. Let's connect next week to discuss that.
Thanks!
rdzhabarov
on 1 Oct 2018
We need to implement ONNX loader for this as well. Reopening.
rdzhabarov
on 12 Oct 2018
Related issues
dati91
路
3Comments
rdzhabarov
路
4Comments
wayneshawn
路
3Comments
ayermolo
路
3Comments
mciprian13
路
4Comments
Most helpful comment
As a followup would be great to have a documentation on how to implement ONNX/C2 operator in Glow.
Specifically, what are the necessary steps and considerations to have to implement an Op.