SysCV / sam-hq

Segment Anything in High Quality [NeurIPS 2023]
https://arxiv.org/abs/2306.01567
Apache License 2.0
3.73k stars 224 forks source link

About interm_embeddings #128

Open zzzyzh opened 7 months ago

zzzyzh commented 7 months ago

As you can see in the code at line https://github.com/SysCV/sam-hq/blob/322488826bda616798901c6280d13a9a90444ae7/train/segment_anything_training/modeling/image_encoder.py#L108C5-L120C36 You have defined a list to store the information for the middle layer, but in the actual application, you are only using the initial layer. May I ask if the rest of the information is of any use? Or did your experiments come up with any relevant structure?