search

hkchengrex / Cutie

[CVPR 2024 Highlight] Putting the Object Back Into Video Object Segmentation
https://hkchengrex.com/Cutie/
MIT License
732 stars 71 forks source link

Weights saving - Best/Last checkpoints ? #93

Closed kassimi98 closed 3 months ago

kassimi98 commented 4 months ago

Hi hkchengrex, Thank you for your work.

We trained Cutie during the pre-training phase for 80,000 iterations and ended up with an exp_id_pretraining_last.pth file. The model saves a .pth file every 10,000 iterations, overwriting the previous one.

I have a couple of questions:

  1. Is the saved pre-trained model the last one or the best one?
  2. How can we evaluate the quality of training in both the pre-training and main-training phases using our own dataset?
  3. If the saved model is the last one, how can we configure it to save the best model instead?
hkchengrex commented 4 months ago

Hi, thanks for the question.

1 -- The last one. 2/3 -- Model selection and validation during training are not implemented here. I'm afraid that you would have to implement your own

kassimi98 commented 4 months ago

Thank you , is there a way to evaluate only after (pre-training / main-training) to see the results of the final model on test data using the same metrics in your paper.

hkchengrex commented 3 months ago

Sure. Just use our evaluation pipeline: https://github.com/hkchengrex/Cutie/blob/main/docs/EVALUATION.md