search

sipeed / MaixPy-v1

MicroPython for K210 RISC-V, let's play with edge AI easier
https://wiki.sipeed.com/maixpy
Other
1.68k stars 438 forks source link

Problem with larger models #337

Open phquanta opened 3 years ago

phquanta commented 3 years ago

I'm trying to get model as large as possible through load_flash API calls, however i've noticed that using Darknet modification of tiny yolo which goes fine through nncase (all of them beta2,beta3 and beta4) with model being 4.4 Mb i'm not able to get inference. Return from flash on sipeed_kpu_err_t returns no error. Return from inference is None. I was playing with C calls in the firmware side, and i've noticed many of the function calls seem to work in nncase v1 RC5 version but not in the v2 such as get model size, get info on layers and etc. I was also wondering if it is possible to get source code of sipeed_kpu_model_load_flash() to see where bug is taking place.

phquanta commented 3 years ago

I think it is related to darkflow due to corrupted ckpt to pb file conversion.

P.S. It is a definetely flaw with Darkflow. Darknet works fine though.

junhuanchen commented 3 years ago

Probably, this part is not open source, but it does not prevent you from using the NNCase interface directly

Neutree commented 3 years ago

you can point out which API you need have bug, we can fix it