Lead Software Engineer (LLMs & Generative AI) @ Indegene | debuggercafe.com | github.com/sovit-123
Added all DINOv3 Transformer (except the 7B backbone one) and ConvNext based Faster RCNN models to Faster RCNN Training Pipeline. All the Transformer based Faster RCNN models use multi-level features from the backbone to build the Faster RCNN head and the ConvNext ones use features from the last layer. Surprisingly, multi-level feature selection for ConvNext Faster RCNN is giving an inferior result compared to the last layer-only feature selection. Maybe need to debug a bit more. Now, just need the compute to pretrain all the adapter heads. Link to the project in the comments.
Lead Software Engineer (LLMs & Generative AI) @ Indegene | debuggercafe.com | github.com/sovit-123
2wProject link => https://github.com/sovit-123/fasterrcnn-pytorch-training-pipeline