search

buptLinfy / ZSE-SBIR

ZSE-SBIR
MIT License
44 stars 7 forks source link

Question regarding tokenization #2

Closed GokulEpiphany closed 11 months ago

GokulEpiphany commented 1 year ago

Thank you for open-sourcing this project. According to the paper, only sketch images are tokenized by Conv2D to increase receptive field. From sa.py, the following line seems to tokenize images and sketches in similar way,

https://github.com/buptLinfy/ZSE-SBIR/blob/e78ee1e0e4ce97e64b0fecb609419466c6f30e50/model/sa.py#L157

Am I missing something obvious?

pyh0715 commented 1 year ago

Learnable Tokenization没开放吧

buptLinfy commented 11 months ago

Thank you for following our research and asking questions!

According to the paper, our network performs best when using learnable tokenizer only on sketches. But for much easier understanding and implementation, sketches and images pass through the same learnable tokenizer in our public code.