mirror of
https://github.com/openai/CLIP.git
synced 2026-11-05 01:37:26 +00:00
[GH-ISSUE #50] YFCC-100M text filtering #47
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
openai/CLIP#47
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @hellbell on GitHub (Mar 3, 2021).
Original GitHub issue: https://github.com/openai/CLIP/issues/50
Hi, thanks for the great work!
I'm collecting YFCC-100M dataset following the paper ("After filtering to keep only images with natural language titles and/or descriptions in English, the dataset shrunk by a factor of 6 to only 15 million photos.").
Could you share the way to filter out the non-natural language texts of YFCC-100M?
@jongwook commented on GitHub (Apr 8, 2021):
We used a third-party language detection library for it. For easier reproducibility, we uploaded the line numbers (the first column of the YFCC100M dataset TSV) corresponding to the filtered subset: https://openaipublic.azureedge.net/clip/yfcc100m_subset_line_numbers.npy.bz2
@hellbell commented on GitHub (Apr 9, 2021):
@jongwook Great, Thanks!
@ModMorph commented on GitHub (Apr 10, 2021):
Hi @jongwook, I can find the YFCC100M dataset images on Amazon commons https://registry.opendata.aws/multimedia-commons/ but the TSV dataset I haven't been able to find. It seems to have been removed from Yahoo, google links go to spammer sites. Any suggestions how to find it?
Are there any other examples of lists of specific parts of public datasaets you can share that CLIP was trained on, along with data that was included/excluded? Trying compare against data that the CLIP model HASN'T seen, and find dataset CLIP has seen and for sure hasn't seen. I guess I can find new images/metadata after x date, but wanted similar datasets so that something like higher camera resolutions for newer data didn't alter my results.
Many thanks, great work!
@jongwook commented on GitHub (Apr 11, 2021):
Hi @ModMorph, I just uploaded yfcc100m.md that points to a TSV file containing the line number, the photo identifier, and the photo hash corresponding to the YFCC100M subset.
It looks like the AWS Open Data website used the photo hash data from the TSV file, which was the third column of the table. They used 2-level prefixes so that a photo whose hash starts with
abcdef...is located underimages/abc/def/.So you can be sure that CLIP has seen these images, and it hasn't seen anything newly taken in 2021.
@ReneTCAd commented on GitHub (Dec 20, 2021):
Hi, I am curious about the specific filtering rule. With the given line number, which text should I use, the title or the description?
@g-luo commented on GitHub (Feb 25, 2022):
@jongwook I was wondering if you had the metadata file for YFCC-100M, which also contains camera id, latitude, longitude and all the hashes? As mentioned above, this data is now difficult to find.
@g-luo commented on GitHub (Feb 25, 2022):
Never mind; found at: https://multimedia-commons.s3-us-west-2.amazonaws.com/tools/etc/yfcc100m_dataset.sql
@dengmengjie commented on GitHub (Nov 27, 2025):
@g-luo
Thanks for sharing the link to the SQL file! I've downloaded it, and it contains the metadata for the Flickr images, but I couldn't find the mapping from the Flickr photo IDs to the MD5 hashes.
I just wanted to double-check, are the hashes indeed missing from the SQL file, or did you find them somewhere else? I'm still looking for the yfcc100m_hash.bz2 file or any other way to map the photo IDs to their corresponding hashes.
Any help would be greatly appreciated!
Thanks again!