Skip to content

Commit d29d25e

Browse files
authoredMay 26, 2024
Add COCO 40k info: GmP CLIP
For reproducing fine-tune with Geometric Parametrization
1 parent 28f040e commit d29d25e

File tree

3 files changed

+162029
-0
lines changed

3 files changed

+162029
-0
lines changed
 

‎COCO/!README.txt

+9
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,9 @@
1+
COCO 40k dataset for testing / reproducing the CLIP GmP implementation: https://huggingface.co/datasets/SPRIGHT-T2I/spright_coco
2+
3+
The 'data' folder with subfolders "0", "1", "2", etc. should be where the json labels are.
4+
Look inside the .json for where the images are expected.
5+
6+
I also just cropped the images to be square (for CLIP's expected input), alas subfolder "data-square" mentioned in default code.
7+
CLIP doesn't appreciate a squished-to-dimensions image, so: Don't forget it! ;-)
8+
9+
The labels in the .json files are capped to fit CLIP's 77 input tokens. You'll just have to download the images above.

‎COCO/short-coco-sprite-train-0_9.json

+144,002
Large diffs are not rendered by default.

‎COCO/short-coco-sprite-val-10_11.json

+18,018
Large diffs are not rendered by default.

0 commit comments

Comments
 (0)
Please sign in to comment.