search

IDEA-Research / detrex

detrex is a research platform for DETR-based object detection, segmentation, pose estimation and other visual recognition tasks.
https://detrex.readthedocs.io/en/latest/
Apache License 2.0
2.01k stars 209 forks source link

Train Dino with focalNet backbone on custom datasets #189

Closed 4BD0U closed 1 year ago

4BD0U commented 1 year ago

I am having difficulty training a Dino model with a FocalNet backbone on a custom dataset. Although I was able to successfully train the model using a SWIN-L backbone, after 10 epochs with FocalNet, all detection scores are still below 1%. I am unable to understand why this might be the case as the dataset used is identical and the pre-trained model provided in the GitHub repository was used as a starting point. Can you help me understand what might be causing this issue, please?

rentainhe commented 1 year ago

I am having difficulty training a Dino model with a FocalNet backbone on a custom dataset. Although I was able to successfully train the model using a SWIN-L backbone, after 10 epochs with FocalNet, all detection scores are still below 1%. I am unable to understand why this might be the case as the dataset used is identical and the pre-trained model provided in the GitHub repository was used as a starting point. Can you help me understand what might be causing this issue, please?

Does this phenomenon occur very often or just a random experiment ？

rentainhe commented 1 year ago

Actually, this error is relatively difficult to locate, and it requires more experiments to observe. If it's convenient, could you provide more information about the custom dataset you're using ?

4BD0U commented 1 year ago

Sorry, it was just an error. My issue was with the Dino repository, not this one. The pre-trained model provided was intended for Dino 5scale with focalNet, but the configuration file corresponded to Dino 4scale.