search

prismformore / Multi-Task-Transformer

Code of ICLR2023 paper "TaskPrompter: Spatial-Channel Multi-Task Prompting for Dense Scene Understanding" and ECCV2022 paper "Inverted Pyramid Multi-task Transformer for Dense Scene Understanding"
MIT License
301 stars 23 forks source link

Channel Task Prompts Wrong #32

Open AaronCIH opened 2 months ago

AaronCIH commented 2 months ago

Hello Authors,

In the Channel Task Prompt Learning Part (code: Multi-Task-Transformer/TaskPrompter/models/transformers/taskprompter.py L249), the line: chan_prompts = self.chan_proj(chan_x) has been comment, which means skipping the channel task prompt results when updates the task prompts. Is that correct? or should I uncomment this line for following paper.

Thank you.

prismformore commented 2 months ago

Hi @AaronCIH ,

It has been a while since I finished this project. One thing I can recall is that on different datasets adding this projection leads to different performance. You may choose to use this projection or not based on your experiment results in your project. For the checkpoints I think they are trained with the current codes.

Thanks a lot.