AbstractPhil/imagenet-clip-features

Update: 10/2/2025

Claude said that I'm not being careful enough with my database curation after grilling me for 20 minutes, so I included the preparer script as well.

Claude Sonnet 4.5 is kind of a chad.

Update; 9/26/2025

Having to download this whole repo is annoying, so I'm making sure the splits are named train/val/test (if they exist) and the named subset is the clip name.

Older non-dated updates

Everything extracted with torch configured as deterministic; using seed 42 on an a100 using colab; so if it has variances from expectation it's on cuda.

It's a little quirky;

• Most of the splits have train, test, val. Many do not.
• Most of the splits have a proper "image_id" md5 id for verification.

The prompts used were direct literal prompts for the classification name;

No use of "a photo of" or any such invariance; just the classification text.

This is a series of clip-vit extracted feature maps from a 256x256 cropped and resized imagenet variant hosted here on huggingface.

I ran the processor 224x224 and then extracted features from the entire dataset batch-sequentially while simultaneously capturing the necessary classifiers and classifications associated with the images for downstream testing and assessment.

Academic and research purpose use only.

clip-vit-large-patch14 variations do exist in the splits.

clip-vit-bigG is the 1280 dim variation and it does exist; it took quite a while to extract - and it is in fact missing it's test split. Sorry about that.

There are many variants of clip-vit-base from many variant forms. Each of them extracted using the same process as the others.