fatih akyon comments

Results 220 comments of


                                            fatih akyon

Nan loss during first epoch in custom dataset training

Started a new training with v8.0.1 :+1:

Nan loss during first epoch in custom dataset training

v8.0.1 gives nan, now trying with v8.0.2 edit: v8.0.4 also gives nan

Nan loss during first epoch in custom dataset training

`cls_loss` increases rapidly at the training start. Then at %1 of training dataset `cls_loss` becomes `inf`. Since the number of target classes is 1, this seems odd. After %10 iterations...

Nan loss during first epoch in custom dataset training

@AyushExel can this be related to the newly introduced losses?

Nan loss during first epoch in custom dataset training

This issue is not present in coco dataset, it is happening when fine-tuning on a custom dataset.

Nan loss during first epoch in custom dataset training

My divergence issue seems to be fixed (10% at 2nd epoch with ultralytics==8.0.17 without nan)

WIP: Add pythonic outputs interface

@AyushExel you can keep it as a generator by removing return and keeping yield. Then implement `__len__()` method of `Result` class so that len(result) would return the true length although...

WIP: Add pythonic outputs interface

I see, then you may have to implement two classes `StreamingResult` and `Result` for iterable and full results. When `stream=True` predict will return `StreamingResult`. You can implement custom `__len__()` method...

WIP: Add pythonic outputs interface

Is there any ETA for this feature? Is it going to be backward compatible with v8.0.6 predict inputs/output? I am asking for my packages that rely on ultrlaytics package :)

All losses went to nan and precision & recall went low & static

@ahthserhsluk what is your imgsz during training?