Sure. I just accessed the dashboard again, and the version seems to be available now. I ran out of credits, which is why it didn’t properly complete all the epochs. If I create a new version now and train from a previous checkpoint (v11), will it work? How many prepaid credits should I buy?
When doing a fine tune, you should get an approximation of how many credits it takes to train. This is typically an overestimate, but we implement that to ensure that you will have enough credits to complete the training. Let me know if that helps!
Based on our current metrics, is it recommended to finish it, or is it better to leave it as is? How much would that improve precision, approximately? Any idea?
Although its unlikely to be the case they would consume the same amount of credits, again this is to be safe and ensure you will have enough credits for training by overestimating a little bit.
This really depends on a large number of factors. From my perspective, you metrics are quite good at this point but it depends on what your KPI targets are. I don’t think your dataset is so big that you have to worry about overfitting, so additional well annotated data wouldn’t hurt. Unfortunately, I’m not exactly sure how much it would improve precision just because of the nature of computer vision and all the confounding factors involved, however well annotated clean data always goes a long way!