Hello @Hamze_Asadi ,
Yes, it is possible that quantization causes your model detection almost random. However, this situation is rare. Min-max range too large causes fl greater than 8. It will affect accuracy.
I advise you to refer aml_npu_sdk/docs/en/NN Tool FAQ (0.5).pdf page 12 title 4.2 to determine it is converting problem or not.
If it is converting exception. You can try to quantize model in int16.
Thank you so much.
right now I am able to generate _nbg_unify folder. there are two things.
1- On page 31 of “Model Transcoding and Running User Guide (1.0).pdf” for ./bin/pegasus export ovxlib
there is option “–postprocess-file POSTPROCESS_FILE”, based on its definition
“”"
a string value to denote the file path of the post-processing configuration file, a .yml file. Multiple tasks can be set in one configuration file. You can either create a post-processing file or use the generation command to generate a post-processing file
“”"
we should have .yml file but I don’t have it, I was wondering where can I get this file. Also is it possible to not have any postprocessing? in my case I just one to get the network output without any further analysis.
2- after generating _nbg_unify folder how can I convert the model to binary so I can run it on the board?
Thank you
Hello @Hamze_Asadi ,
postprocess-file is not necessary. You do not need to take it. After running ./bin/pegasus export ovxlib successful you will find a file named vnn_post_process.c in _nbg_unify folder. You can refer this file about how to decode result to float32. You can also refer our demos to do it.
If converting succeed, the model which ends with .nb will generate in _nbg_unify floder. It is the model that can run on the board.
Sorry, I think I didn’t ask my questions clearly,
After transcoding I will get a network.nb file which is a binary file of the network. with some other post and pre processing c files. Right now my question is how to run this network.nb file on the board I couldn’t find any instructions?
Thank you for your guide.
Hello @Hamze_Asadi ,
Here is document about how to run on the board.
aml_npu_sdk/docs/en/DDK_6.4.8.7_SDK API (1.9.4).pdf at master · khadas/aml_npu_sdk · GitHub
