search

frank-xwang / UnSAM

[NeurIPS 2024] Code release for "Segment Anything without Supervision"
403 stars 27 forks source link

Custom training #13

Open NicDionne opened 4 weeks ago

NicDionne commented 4 weeks ago

Hi, your results are really interesting, and we want to try the method on our own dataset to see if it outperforms SAM.

However, I have a few questions regarding the training process, specifically about the train.sh script. Some guidance would be greatly appreciated on the following points:

Custom Dataset for Unsupervised Part In the script, I see the following line: export DETECTRON2_DATASETS=/PATH/TO/YOUR/DATASETS Is this where I should place my custom dataset for the unsupervised part? I'm a bit confused because, according to this README, I need to place the dataset in the tsv directory. Isn’t that in a supervised format?

Supervised Dataset Another line in the script is: export TRAIN_DATASETS=/PATH/TO/YOUR/SA-1B Is this where I should place my supervised dataset in the SA-1B format?

Running the Script Is train.sh the only script I need to run to obtain a custom model?

UnSAM+ Model Lastly, how can I go about obtaining a UnSAM+ model?

wenfang361 commented 1 week ago

Hi, Have you made any progress? Has the issue been resolved? I've encountered a similar problem.

NicDionne commented 5 days ago

Unfortunately, the issue has not been resolved.