search

facebookresearch / EgoVLPv2

Code release for "EgoVLPv2: Egocentric Video-Language Pre-training with Fusion in the Backbone" [ICCV, 2023]
MIT License
85 stars 11 forks source link

Pre-extracted EgoNLQ Features seem unavailable #15

Open StOnEGiggity opened 5 months ago

StOnEGiggity commented 5 months ago

Hi, thanks for your work. I downloaded Pre-extracted EgoNLQ Features multiple times and failed when downloading 50% each time. Do you have any idea about that? Thank you.

ShramanPramanick commented 5 months ago

Hi, Thanks for your interest in our work. I just tried to download the tgz file with EgoNLQ pre-extracted features, and it worked fine (as described here).

image

The tgz file requires ~40GB space. Will you please ensure if you have this much space available on your disk? Even if that does not work, I can upload the file into a drive and share it with you.

StOnEGiggity commented 5 months ago

Thanks for your reply. I guess it may be related to the regional network (in Australia for now). It is better to provide a drive link. Thank you very much.

StOnEGiggity commented 5 months ago

By the way, I am trying to extract features with the provided pre-trained model, I found some weight of Roberta Model were not initialized from the model checkpoint, such as "roberta.encoder.layer.10.alpha_t2i" and "roberta.encoder.layer.10.crossattention_t2i.output.dense.weight". Do you have any idea about this? Thank you.

ShramanPramanick commented 4 months ago

I have now uploaded the EgoNLQ features into OneDrive. Please find them here. Please let me know if you can download it now.

Regarding your second question about extracting features yourself, please clarify which script you are running and show me your exact Python command so I can better understand and correctly assist you. Thanks!

StOnEGiggity commented 4 months ago

Thank you very much~ I checked the code and found it is from the code ".from_pretrained". So it looks like no problem for now.