Can tiny-yolo-v3-coco train on its own data set?

[ystyle]

Can tiny-yolo-v3-coco train on its own data set?

[chewxy]

I don't understand the question. Do you mean if you don't want to use pjreddie's weights, and want to train your own?

Yes you can do that. @LdDl wrote the model in such a way that it optimizes for the fwd only operations, but you can still get a graph out of it, and then filter for input nodes

[chewxy]

here's a tip:

func FilterVar(ns G.Nodes) (retVal G.Nodes){
    for _, n := range ns{
        if n.IsVar(){
            retVal = append(retVal, n)
        }
   }
 return retVal
}

Then you can just call tinyyolo.g.AllNodes() and pass the result into the Filtler func to get the results that you can use in a Solver.

You gotta write the cost thing yourself tho.

Eager to help if you want to submit a PR.

[LdDl]

@ystyle
Working on training mode for YOLOv3 tiny in this fork... Losses are going down while training. But I'm not doing PR to main Gorgonia's repository due I have 2 blocks:

1. I have not much time currently due my daily job
2. Checking correctness
   I'm trying to check it "by hand" (looking at numerical output then check it in AlexeyAB's fork). I do not have CUDA-card on my home PC (broken), so can't train anything in 1 hour at least and check if losses are dropped down.
   There are some suspicious things such as https://github.com/gorgonia/tensor/blob/master/ap.go#L280 also. Sad thing, that I can't help to guys with those due reason # 1 :(

+1 for PR, if you know "stuff" around YOLO's losses.

[dcu]

development was moved here https://github.com/LdDl/yolo-go right?

[LdDl]

development was moved here https://github.com/LdDl/yolo-go right?

@dcu
Yes, it did, currently.
I stuck a bit there with errors estimation.