designsystem-set-membership-siglip2-demo

Design System Set-Membership SigLIP2 Demo Checkpoint

This repository contains the trained checkpoint and local config/processor files used by the Design System set-membership web demo.

Files

• checkpoint.pt: PyTorch checkpoint for the set-membership discriminator.
• hf/google--siglip2-base-patch16-224/: local Hugging Face config/processor files for google/siglip2-base-patch16-224.

The demo app instantiates the SigLIP2 architecture from config and then loads the trained checkpoint weights from checkpoint.pt.

Checkpoint Provenance

Source path on Gautschi:

/scratch/gautschi/jasonwu/designsystem/design_discriminator_runs/setmem-default-additive-x-res-bce-pat4000-20260630T053443Z/checkpoint.pt

SHA-256:

8eeb0b943e7bbb8cb30dcf2ca2302b3ab50bc853c67e4f35165fd26bbfc87092

Training args subset:

model_name = google/siglip2-base-patch16-224
compatibility_mode = additive_x_residual
set_encoder = transformer
set_hidden_dim = 512
direct_scorer_layers = 2
preprocess_crop_top = 0.1
preprocess_crop_bottom = 0.1
preprocess_target_short_side = 256

Intended Use

This checkpoint is intended for the companion Flask web demo package. It scores whether a candidate screenshot belongs to the same design system as uploaded context screenshots.

Uploaded images should be preprocessed before SigLIP encoding with the checkpoint training settings: crop 10% top, crop 10% bottom, resize short side to 256.