Sergiy

> @Robotatron Have you tried it? I think it will give the AP for [email protected]. I figured it out. It seems my colleague modified the coco eval code and he...

Reproducible with `yolo train segment mosaic=0.5`

Your losses are huge, the net probably can't predict anything useful, hence the APs are 0. My segm loss starts around 2 from pre-trained model, your loss is 10

Try to do inference on your model with a very low confidence threshold and see what the model predicts

Similar issue https://github.com/ultralytics/ultralytics/issues/508 Seems like a general bug?

> @ZicsX @Robotatron save_dir is not a valid key. To save to PROJECT/NAME directory: > > ``` > yolo train project=PROJECT name=NAME > ``` Yes, it's not a valid key...

do you have image rotation augmentation enabled? Could be that

So do the bboxes break only with mosaic? Do bboxes are drawn correctly without mosaic augmentation? Otherwise check your data in FiftyOne to see if bboxes are correct in the...

> This just allows the code to run but is not a fix. If it runs with the desired output, why is not a fix in your opinion? I assume...

Just trained a model without any augmentations at all (with `--hyp hyp.no-augmentation.yaml`) and it still crashes. What's weird is that everything worked fine 2 days ago and did not made...