rknn.build failure at Global Average Pool input height/width > 128 about rknn-toolkit2 HOT 11 OPEN

Global avgpool input height/width support > 128 now for new version 1.2.0, but still should < 343

from rknn-toolkit2.

@rkchrome

Thank you so much for your update. I tried it and it looks like it's making a big progress on my test model. However, this time got stuck at "Fatal: Meet unsupported operator: Sqrt(name="Sqrt_941")."

The error messages as below:

(rknn2) paul@tensor:~/pytorch/AdaBins$ python rk3566onnx.py 
right after import torch....
W __init__: rknn-toolkit2 version: 1.2.0-f7bb160f
--> config model
done
--> Loading model: ./rknn_models/model_3566.onnx
done
--> Building model
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
W build: Global Average Pool input height/width > 343, which is not suppoted!
Analysing :   0%|                                                           | 0/
Analysing :  34%|███████████████▉                               | 262/770 [00:00
Analysing :  52%|████████████████████████▍                      | 400/770 [00:00
Analysing :  65%|██████████████████████████████▎                | 497/770 [00:00
Analysing :  77%|████████████████████████████████████▊           | 590/770 [00:0
Analysing : 100%|███████████████████████████████████████████████| 770/770 [00:00<00:00, 1067.81it/s]
I RKNN: librknnc version: 1.2.0 (b162fe95e@2022-01-14T14:12:38)
D RKNN: [11:23:26.989] RKNN is invoked
Fatal: Meet unsupported operator: Sqrt(name="Sqrt_941").

It seems that it cannot build because rknn does not support "Sqrt" operator.

Please advise how to fix this issue. Thanks a lot.

from rknn-toolkit2.

Sqrt is not supported temporarily

from rknn-toolkit2.

Do we know when Sqrt well be supported? Thank you.

from rknn-toolkit2.

Do we know when Sqrt well be supported? Thank you.

Sqrt may be supported on next format release version as soon.
If you need Sqrt in LayerNorm, maybe next version will open graph-optimization for it.

from rknn-toolkit2.

@HunterRockchips,

I upgrade to the latest version 1.3.0 and indeed the Sqrt error is gone! Thank you very much.

But now I encounter the following error:

Fatal: Meet unsupported operator: CumSum(name="CumSum_1274").

When would the CumSum operator would be supported? or What command I can change in my code to avoid this unsupported error?

Thanks a lot for your help.

from rknn-toolkit2.

This op(CumSum) is not in develop plan. I suggest you change you graph with other operators which had supported.

from rknn-toolkit2.

@HunterRockchips,

I upgrade to the latest version 1.3.0 and indeed the Sqrt error is gone! Thank you very much.

But now I encounter the following error:

Fatal: Meet unsupported operator: CumSum(name="CumSum_1274").

When would the CumSum operator would be supported? or What command I can change in my code to avoid this unsupported error?

Thanks a lot for your help.

meet the same question

I rknn building ...
I RKNN: librknnc version: 1.3.0 (c193be371@2022-05-04T20:23:58)
D RKNN: [10:47:53.678] RKNN is invoked
Fatal: Meet unsupported operator: Sqrt(name="Sqrt_51").

to:rockchip
sqrt not supportin rknn2 1.3version,please add the op in it next version

from rknn-toolkit2.

@HunterRockchips

After I replace torch.cumsum with equivalent python code in my model, now it encounters a new unsupported operator:
Fatal: Meet unsupported operator: ScatterND(name="ScatterND_1308").
So are we going to add "ScatterND" op soon?
Thank you.

from rknn-toolkit2.

Temporarily not. Maybe Sqrt will support quicker.

from rknn-toolkit2.

Or, you can cut the model into 2 pieces, and execute this op by cpu.

from rknn-toolkit2.