U2PL icon indicating copy to clipboard operation
U2PL copied to clipboard

Published 20 hours ago • verified publisher icon Haochen-Wang409

Train custom data

Open Holmes2002 opened this issue 2 years ago • 3 comments

Can you explain how to train custom data please ?

Holmes2002 avatar Aug 06 '22 02:08 Holmes2002

Just modify data-specific parameters (e.g., data_root, data_list num_classes) in config.yaml.

Haochen-Wang409 avatar Aug 09 '22 13:08 Haochen-Wang409

Should I adjust parameter num_negatives in config file for custom data ?

Holmes2002 avatar Aug 12 '22 03:08 Holmes2002

Also, how to generate SegmentationClassAug for custom data? Any tips on this?

LuizZillmann avatar Aug 12 '22 03:08 LuizZillmann

Should I adjust parameter num_negatives in config file for custom data ?

The num_negatives means the number of negative keys per query in contrastive learning. Our method is robust against this parameter. Thus, I don't think it worth tuning for custom data. However, if you fail to get expected results, it can be ablated on your dataset.

Haochen-Wang409 avatar Aug 16 '22 05:08 Haochen-Wang409

Also, how to generate SegmentationClassAug for custom data? Any tips on this?

SegmentationClassAug is the ground-truths for each input images. Assume we have an image with shape HxWx3, thus the corresponding ground-truth should be HxW. And each emelent denotes the class for each pixel.

Haochen-Wang409 avatar Aug 16 '22 05:08 Haochen-Wang409