These are chat archives for beniz/deepdetect

24th
May 2017
alkollo
@alkollo
May 24 2017 14:07

Hello,

Would be fine to have loss displayed for caffe backend like on the original caffe doc :

I1203 solver.cpp:204] Iteration 100, lr = 0.00992565
I1203 solver.cpp:66] Iteration 100, loss = 0.26044

instead of

INFO - 14:03:36 - Iteration 640, lr = 0.01
INFO - 14:04:03 - Iteration 680, lr = 0.01

I did not see any info about tuning this in deepdetect config.
cheers,

Emmanuel Benazera
@beniz
May 24 2017 15:15
loss is from the API with all other metrics.
Emmanuel Benazera
@beniz
May 24 2017 16:51
I'll check why the train loss may not show up though
alkollo
@alkollo
May 24 2017 17:18
thx, I know about the API call but I alway have an eye on the server output, would be usefull anyway to have a glance at the loss to know if model is training without a call.
Emmanuel Benazera
@beniz
May 24 2017 17:19
OK
alkollo
@alkollo
May 24 2017 17:19
Anyway why inception > 1 are not available to train ? something impossible to adapt ?
Emmanuel Benazera
@beniz
May 24 2017 17:32
I dont think I ve seen it with caffe. Personally I likebtthe resnet 50.
alkollo
@alkollo
May 24 2017 17:37
ok thanks. What crop size do you use with resnet50 ?
Emmanuel Benazera
@beniz
May 24 2017 17:43
for many models, I don't use cropping but mirroring and the noise parameters.