search

SunzeY / AlphaCLIP

[CVPR 2024] Alpha-CLIP: A CLIP Model Focusing on Wherever You Want
https://aleafy.github.io/alpha-clip
Apache License 2.0
653 stars 38 forks source link

Guidance needed: Processing GRIT-20M dataset in .parquet format for Alpha-CLIP #60

Open qingpowuwu opened 1 month ago

qingpowuwu commented 1 month ago

Hello,

I'm working with the GRIT-20M dataset for the Alpha-CLIP project as described in the training README . However, I've encountered some discrepancies between the instructions and the dataset format I've obtained.

  1. Dataset Format:
    • The data preparation script (sam_grit.py) is configured to use .tar files, as evidenced by the line: parser.add_argument('--tar-pth', type=str, default="GRIT-1m/00001.tar")
    • However, the dataset I've downloaded is in .parquet format (e.g., coyo_0_snappy.parquet, coyo_10_snappy.parquet, etc.).
    • Could you confirm if this .parquet format is correct for the latest version of the dataset?

Thank you for your time and assistance.

SunzeY commented 1 month ago

you can follow the download script in KOSMOS-2 to download .tar file. If you download from hugging face, you need to adjust the script. (by the way, this script only use SAM to change box into masks, its easy to reimplement it in .parquet format)