What is the use for average_loss
? Would someone please give an example of it or explain it in layman's terms?
Asked
Active
Viewed 4,645 times
1 Answers
8
You can check in the caffe.proto file. Line 151 in the present version gives the following comment for average_loss:
Display the loss averaged over the last average_loss iterations

malreddysid
- 1,342
- 1
- 9
- 12
-
1so what does that mean exactly? what does averaging loss for x iteration do? just averaging and displaying it ? why would I want that? – Hossein Oct 22 '16 at 13:54
-
Instead of the complete average loss. It gives you the loss for the last n iterations. It's just a feature. Might be useful in some networks. – malreddysid Oct 22 '16 at 21:49
-
Thanks, though it would be very desirable to know an intuitive example for it – Hossein Oct 23 '16 at 15:59
-
8Sometimes training is performed with a small batch size and the loss might be noisy. It makes sense to average it over last n iterations to get a better understanding of what is happening. – warmspringwinds Jan 13 '17 at 23:16
-
Thanks a lot, that makes sense – Hossein Jan 27 '17 at 10:43