-
Dear all, We (@isabelperezf and me) are using Sleap in the CLI for labeling frames and Colab for training but have some doubts: We have already trained a top-down model in Colab, downloaded the "models" folder, opened the project locally, ran predict using the downloaded .json files, corrected the worst instances (according to the score). Now our doubts start: Thanks! A |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
Hi @auesro and @isabelperezf, You are correct that training does not pick up from the pre-trained model. This is intentional as we find that randomly initialized models trained from scratch every time are more stable and can train just as well. This is partially because we avoid "reinforcing" errors that would be learned early on when there is little labeled data available and the model is heavily overfitting to it. That said, if you want to try it yourself, we have a notebook on resumable training you're welcome to run through. It's a bit more advanced than the basic training notebook, but if you follow it step-by-step it'll do what you're looking for. Let us know if you this works for you or if you run into any issues. |
Beta Was this translation helpful? Give feedback.
-
Dear @talmo Thanks! A |
Beta Was this translation helpful? Give feedback.
Hi @auesro and @isabelperezf,
You are correct that training does not pick up from the pre-trained model. This is intentional as we find that randomly initialized models trained from scratch every time are more stable and can train just as well. This is partially because we avoid "reinforcing" errors that would be learned early on when there is little labeled data available and the model is heavily overfitting to it.
That said, if you want to try it yourself, we have a notebook on resumable training you're welcome to run through. It's a bit more advanced than the basic training notebook, but if you follow it step-by-step it'll do what you're looking for.
Let us know if you this works for y…