dandelin / ViLT

Code for the ICML 2021 (long talk) paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision"
Apache License 2.0
1.36k stars 209 forks source link

Question about pad_choice #53

Open Richar-Du opened 2 years ago

Richar-Du commented 2 years ago

https://github.com/dandelin/ViLT/blob/762fd3975c180db6fc88f577cf39549983fa373a/vilt/modules/vision_transformer.py#L636

Thanks for your awesome work! I have a question about pad_choice. When choosing pads to pad the sequence, you sample invalid patches with replacement=True. I wonder if this will bring repeat pad patches to the sequence?

Thank you :)