Open Bergylta opened 2 weeks ago
Is my following understanding correct? I think that you are trying to retrain a model with Paulina's data, by taking an already existing baseline model and now retraining it. Is that correct?
From the following message: _train: WARNING ⚠️ /mimer/NOBACKUP/groups/snic2021-6-9/tmp_dir/KSO_Paulinaframes/images/01211c55-4dd6-4453-981a-2bc85bbbbb18.jpg: ignoring corrupt image/label: non-normalized or out of bounds coordinates [] I understand it as that the data from Paulina does not fit the model. It says that the data from Paulina is 'non-normalized or out of bounds'. So I guess (but not sure) that that means that the values for every pixel in the image does not match the model. For example, that the model was trained on data that have pixel values from 1-255 and that Paulina's data goes from 0-1 for example. Or something like that. But I am not sure about this as I am not that familiar with the training process and the specific model. Does this ring any bells and that you know how to trouble shoot? Or do you want me to further investigate?
@Diewertje11 In this case, we are taking a baseline model (yolov8 baseline segmentation) and using it to train on paulinas data that she has annotated. I think you are correct in that the pixels might not match, there was something that happened when zooniverse did a update a while back, and that our bounding boxes were not centred above the object, but instead was a bit off, which might be related to this (also unsure). But this is a new issue that i haven't seen before i think, nor do i know how to troubleshoot it sadly. We can put this at prio 2-3
🐛 Bug
The workaround might have worked, but a new issue showed up in the train_models.ipynb notebook, unsure if this has to do with the baseline model, Alvis, SNIC, mlflow or something else entirely.
To Reproduce (REQUIRED)
Batch size: 8 Epochs: 50 mlp.registry= mlflow Model: YoloV8 Segmentation Input:
Output:
Additional context
Add any other context about the problem here.