search

ceccocats / tkDNN

Deep neural network library and toolkit to do high performace inference on NVIDIA jetson platforms
GNU General Public License v2.0
718 stars 209 forks source link

different BN epsilon from the forked darknet #167

Closed thnkinbtfly closed 2 years ago

thnkinbtfly commented 3 years ago

Hi, thanks for this great project!!

Some outputs from the forked darknet and the tkDNN was different, and I tried to find out the reason.

I think the default epsilon used in tkDNN (1e-5) is different from darknet (1e-6).

The forked darknet uses 1e-6, as below. https://git.hipert.unimore.it/fgatti/darknet/blob/master/src/blas.c#L292

But tkDNN is using 1e-5. https://github.com/ceccocats/tkDNN/blob/b8855b9599e52a51b371e99255063cd6f00fecd7/include/tkDNN/Layer.h#L35

After updating this, the difference I confronted was resolved! Could you check this?

thnkinbtfly commented 3 years ago

I just checked the original darknet changed the epsilon to 1e-5. https://github.com/AlexeyAB/darknet/commit/a34682e67c9ff279019e17070a5d34d596cd8422

Could you consider using the same epsilon when exporting the outputs and running the network?

mive93 commented 3 years ago

Hi @thnkinbtfly I don't get the point. So you're aying that he's using 1e-6 or 1e-5? And where do you experienced the error?

mive93 commented 2 years ago

Closing for inactivity. Feel free to reopen.