Nvidia digits to train cnn with caffe

Dear all,

Do you know if it is possible to use Nvidia digits software to train CNN and then convert it to openmv ?


Hi, yes and I used it before.

Nice, do you have some advice because i dont manage to quantize to openmv camera with :

python2 ~/openmv/ml/cmsisnn/nn_quantizer.py --gpu --model ./quantizer.prototxt --weights ./snapshot_iter_240.caffemodel --save ./V22_Lenet.pkl

I used a lenet network with my own image set.
As nvidia digits is used in a docker, i modified the directory in the generated prototxt but it stops at this step :

I0404 11:27:20.907341 24927 net.cpp:257] Network initialization done.
I0404 11:27:20.940609 24927 net.cpp:746] Ignoring source layer train-data
F0404 11:27:20.940631 24927 blob.cpp:496] Check failed: count_ == proto.data_size() (500 vs. 0) 
*** Check failure stack trace: ***
Abandon (core dumped)

I joined the original and modified prototxt files.

I tried with a custom network from an openmv network but with the same results …
I should misunderstand a step. Thanks for your help,
nvidiaDigits_files.zip (1.5 KB)

Maybe it’s a different version of caffe, I think you need to install it from source: