BN error

[fzh-adham]

hi when I want to generate c files I get this error "the checkpoint file contains 1-dimensional weights for conv1.bn.weight with dimensions (24,). Ensure the BatchNorm layers have been folded." what is the problem and what is the solution?

[ermanok]

Hello,

The hardware does not support BatchNorm (BN) layer so if you want to use a BN layer in your model, you have to fuse these layers to the preceding convolution layers (see here). If you enable Quantization Aware Training (QAT) during your model training, the code automatically folds the the batchnorm layers. So if you use qat_checkpoint.pth.tar or qat_best.path.tar, you will not get such an error during the code synthesis. If you want to use post training quantization without QAT, you may run batchnormfuser.py to fold the BN layers to the convolutional layers.

[fzh-adham]

Hi Thanks for reply Yes,I have trained in both ways and I just have problem in none qat So for your second choice in command how do I run my model in training with batchnormfuser.py Could you please tell me in details🙏🙏

On Thu, 25 Jul 2024, 12:35 Erman Okman, @.***> wrote:

Hello,

The hardware does not support BatchNorm (BN) layer so if you want to use a BN layer in your model, you have to fuse these layers to the preceding convolution layers (see here https://github.com/analogdevicesinc/ai8x-training?tab=readme-ov-file#batch-normalization). If you enable Quantization Aware Training (QAT) during your model training, the code automatically folds the the batchnorm layers. So if you use qat_checkpoint.pth.tar or qat_best.path.tar, you will not get such an error during the code synthesis. If you want to use post training quantization without QAT, you may run batchnormfuser.py to fold the BN layers to the convolutional layers.

— Reply to this email directly, view it on GitHub https://github.com/analogdevicesinc/ai8x-synthesis/issues/350#issuecomment-2249838414, or unsubscribe https://github.com/notifications/unsubscribe-auth/AWLZTJMP5HZEYDLOCBTZ2VTZOC5TZAVCNFSM6AAAAABLMOLD62VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENBZHAZTQNBRGQ . You are receiving this because you authored the thread.Message ID: @.***>

[ermanok]

It is a python script that you can run from the command line. The details of the batchnormfuser is here. But we always suggest to use QAT as the direct quantization of a floating model may cause huge accuracy drop.

[fzh-adham]

Thanks a lot

I have another problem in command

In the documentation is suggested to use --energy for energy during processing but in my max78000fthr board it doesn't show

How can I get energy ?

On Thu, 25 Jul 2024, 13:46 Erman Okman, @.***> wrote:

It is a python script that you can run from the command line. The details of the batchnormfuser is here https://github.com/analogdevicesinc/ai8x-training?tab=readme-ov-file#command-line-arguments-1. But we always suggest to use QAT as the direct quantization of a floating model may cause huge accuracy drop.

— Reply to this email directly, view it on GitHub https://github.com/analogdevicesinc/ai8x-synthesis/issues/350#issuecomment-2249979424, or unsubscribe https://github.com/notifications/unsubscribe-auth/AWLZTJJ4F5JSUTFC4S7KJWTZODGALAVCNFSM6AAAAABLMOLD62VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENBZHE3TSNBSGQ . You are receiving this because you authored the thread.Message ID: @.***>

[ermanok]

You need a MAX78000 EvKit to measure and observe the energy consumption. The feather board does not have an energy measurement circuitry.

[fzh-adham]

thanks for reply is there any way to calculate energy consumption and power?

[ermanok]

There is no other straightforward approach to measure the energy consumption.

[github-actions[bot]]

This issue has been marked stale because it has been open for over 30 days with no activity. It will be closed automatically in 10 days unless a comment is added or the "Stale" label is removed.