-
Notifications
You must be signed in to change notification settings - Fork 127
Missing 'InstanceNormalization' operator #18
Comments
Any idea if/when this will be added? It would be really useful for a project I'm currently working on. |
Hi, I would love to take on this issue. Is anyone else working on this already? |
@gnsmrky and @JamesOscar - Thanks for your interest in ONNX.JS. #82 should have the IN operator for the @yhung119 - I have a simplistic implementation for this already. Please feel free to take a look to see if it can be improved. Additionally, the IN operator for the |
@hariharans29 #82 has few errors. Is it still good to use it? Or should wait until official stable release? I actually worked around this issue by implementing InstanceNorm in PyTorch and managed to run it on ONNX.js now. |
Thanks for your comment. It's great to know that you have a work-around to unblock yourself. Some comments/questions -
|
@hariharans29, thanks a lot! The background info is that I re-wrote the InstanceNorm using basic ops. It turned out the performance is not that bad as most basic ops are
|
Closing this as #82 is merged to master and primarily addresses the issue of missing IN op in ONNX.js. Currently, cpu and wasm backends are supporting IN op. |
Lately Instance Normalization (IN) has become popular. Just curious if IN is being planned to be added? Would love to see that for wasm backend.
ONNX.js operator list:
https://github.com/Microsoft/onnxjs/blob/master/docs/operators.m
BTW, ONNX runtime does have IN supported:
https://github.com/onnx/onnx/blob/master/docs/Operators.md#InstanceNormalization
The text was updated successfully, but these errors were encountered: