How to get all object super link of $$H_{link}$$ in last-layer hidden states during testing ? #14

Activity

1. josephzpng commented on Jan 8, 2025

   

   josephzpng

   on Jan 8, 2025

   More actions

   I have read issue 11, but still have some questions. During testing, you utilize the forward function for testing

   VisionLLM/VisionLLMv2/visionllmv2/eval/eval_det.py

   Line 119 in 028f8b3

   result = model.forward(**data, return_dict=True)

   But as far as I know, the forward function cannot perform the next token generation, resulting in struggling to output answer and [DET] token. Why don't you use the generate function? Is it possible that the inputs ids in your test contain the answer?

   When will you make the training and testing datasets and dataloader public?

   VisionLLM/VisionLLMv2/visionllmv2/eval/eval_det.py

   Line 111 in 028f8b3

   dataset = eval_dataloader.dataset

   I have the same question, do you have any new insights?