Fine tuning
Opened this issue · 11 comments
Hey I have been fine tuning using pre-trained checkpoint . however Having issues for it to converge , I have 20k Mask and image set for background removal . I have run 30 Epochs and the results are worse than the pretrained one
Any advice on how many epoch or learning rate ?
Finetuning with additional data only may cause catastrophic forgetting. Try adding your data to the composite dataset.
Finetuning with additional data only may cause catastrophic forgetting. Try adding your data to the composite dataset.
I'm just curious - what's the reason of that?
So you mean i Need to train on the pretraining dataset and my own dataset ? Can this be an issue with my dataset?
@plemeri I saw that in your transparent background libraries nightly model scores much higher than HR model in model zoo when tested on DUTS-TE . Can you share the reason and training technique ? is it just trained longer ?
Nightly model was trained with the updated training strategy which was my doctorate dissertation. Currently I have no time to release updated contents.
can you share that dissertation document ? i can update the code
@plemeri if you can share it
My dissertation is not publicly opened yet because of the ongoing patent registration. Sorry for that.
@plemeri congrats on your doctorate dissertation! 👍
I would be more than happy to see the updates on this repository and not kill as everyone does after some time. We are waiting updated training strategy at some point.
Hello all, just for the heads up, the patent office just started to apply the patent to the US. However, I'm afraid there are better methods out there since our work has been released, so it would be better using their work but I'll release the updates as soon as the patent process finishes. Until then!
Could you point us in the direction of these better methods then? :)