How my network can be convergent quickly ? about darknet HOT 9 CLOSED

After 900 rounds training I found obj value still < 0.00xxx, is there any thing wrong ? Could you help us about this slow convergent issue ? @sowson

I have tried learn rate 0.001 and 0.0001 to train network, but the same ...
The avg loss like 11.xxx -> 10.xxx -> 12.xxx -> 9.xxx -> 12.xxx -> 13.xxx -> 10.xxx -> 10.xxx -> 11.xxx by time, so I have to request help from you please.

from darknet.

I need to think about it more... the thing is that trained CNN works even if loss not showing this. I compared many times original project and this fork but do not see any issue with the loss calculation... Thanks for this challenge, I try to help you as well. Please give me some time to figured it out.

from darknet.

Thank you so much @sowson, I expect your feedback any time.
I tried:
learning rate 0.001, 0.0001
batch size and sups: 64/8, 64/16, 32/8
burnin: 0
random:1, 0
anchors: voc default, from my kmean
....
But any combination works bad based on your framework, all the avg loss just can down to about 10 when train 1000 rounds. In addition my training loss will get to 10 about 30 rounds, then it will up to 9 then down to 12, then up to 10 then down to 11 all the time, so this is my case, help me please

from darknet.

hi @sowson, is there any update for your research?

I think there are 5 factors for our training (0. training hardware [we are different but not key factor to final avg loss]. 1. training framework like your project [we are same]. 2. training dataset [we can train based on same nfpa dataset I involved before]. 3. training yolov2.cfg conf file [we are different please share with me about your conf file in order we can keep same]. 4. training rounds and progress log [can you share your log about some big change part for example which round your loss value change to < 50 ? which round change to < 10 ? < 5 ? < 1.0 ? < 0.1 ?

so could you please upload your yolov2.cfg conf file which i can compare with mine. And share some key info what I mentioned as above? Thanks so much Sowson.

from darknet.

Here you go...
yolov2-n.cfg.txt
Thanks!

from darknet.

thanks so much sowson.
How about your training progress and some change checkpoints as I mentioned above ?

from darknet.

@sowson I have seen your yolov2.txt, but I found maybe [region]
anchors = 6,14, 70,82, 176,190, 291,375, 382,377 is not expected, and I have searched articles about anchors in yolov2 which suggest to value from 0.0 to 13.0, but yours is not, so can you train your data to get good result ? I just double confirm with you about my confuse. Thank you.

from darknet.

This is default yolov2-voc.cfg from the original author and this repository as well. I only changed
from:
learning_rate=0.001
burn_in=1000
to:
learning_rate=0.0001
burn_in=100
Of course, on top of the file, training part and testing part is switched to train.
Thanks, and enjoy!

from darknet.

Fix, Thanks so much sowson !

from darknet.