CODEWITHSUNDEEP
image-processingtensorflowconv-neural-networkobject-detection
Odgiiv
Odgiiv

Reputation: 713

Google Object Detection API: Fluctuation in TotalLoss

I am using Google Object Detection API with my own dataset. Mostly after 50K steps it begins to converge with 60 percents accuracy. I think it works fine in general. But when if you look at TotalLoss graphic or in general all loss graphics, it fluctuates so much. It looks like this:

enter image description here

What could be the reason for this? Is it normal or not? If not what is explanation. Also occasionally, I see in the example images some bounding boxes are doubled in one area, why is that?

Upvotes: 0

Views: 735

Answers (2)

yagmur
yagmur

Reputation: 323

yes as @Jonathan explained, 'wiggles' are mostly observed when you have smaller batch sizes. Unfortunately, you are going to need at least 16GB memory to handle batches more than 1 when using Faster-RCNN on Tensorflow API. The only way except you don't have much power of processing is just to be patient until some thousands of iterations completed, in my case it was more than 100,000.

Upvotes: 0

Jonathan Huang
Jonathan Huang

Reputation: 1558

Yes, fluctuation in the loss is very normal particularly because the detection pipelines are usually trained with small batch sizes (batch size 1 in the case of Faster R-CNN), so you typically only see a meaningful decrease in the loss if you average over many steps.

Upvotes: 3

Related Questions