ACS on MedMNIST: checkpoints, part 3
Third part of the trained models behind the ACS paper. Part 1 (main table, G1/G2, E5, E7) is xulala/acs-medmnist-all-checkpoints; part 2 (J2, J3, NCF controls, OrganC/Pneumonia, EyePACS, docs/59 herding-vs-random, and the part-1 overflow that didn't fit under HF's 20,000-file cap) is xulala/acs-medmnist-checkpoints-extra. This third repo holds the two experiments that fit in neither:
| experiment | cells | what |
|---|---|---|
prereg40 |
375 | Within-dataset per-class-budget sweep (4/8/13/20/40) on Retina/Breast/Derma/OCT/Tissue, three selectors (random/herding/coverage), to test whether coverage's small-budget failure is budget-driven |
acf56 |
174 | Appearance-consistency filter (ACF) intervention: does filtering by pixel-space label consistency fix coverage's small-budget failure (plus random-deletion and reverse-direction controls) |
Same training protocol (ResNet-18 from scratch, 224 px, 1000 epochs, no augmentation, deterministic,
final-epoch checkpoint) and the same layout as parts 1-2:
<experiment>/<dataset>/r<ratio>/<arm>/seed_<s>/ with final.pt, selected_indices.npy,
predictions_clean.npz, metrics.jsonl, run_config.json, training_history.json.
manifest_part3.csv: every cell with test BA, worst-class recall, checkpoint sha256 (recomputed
before upload and checked against the value logged at training time) and selection sha256.
Numbers and references audit
Every metric quoted in the paper is checked against the result files in this repo's companion GitHub repo, and every citation's verification status is disclosed (most carried over from an earlier-reviewed draft; a few added later and explicitly marked unverified, since web access was unavailable when they were checked): REFERENCES_AND_NUMBERS_AUDIT.md.
J4 and J6 (added 2026-10-19)
| experiment | cells | what |
|---|---|---|
j4aug |
350 | J4 (docs/53): main-table selections retrained WITH data augmentation |
j4pt |
350 | J4 (docs/53): augmentation + ImageNet-pretrained ResNet-18 (j4_regime.json in each cell) |
j6 |
240 | J6 (docs/54): ACS and the Graph-A2 objective rebuilt on DINOv2 / CLIP features (arms j6acs_<enc>, j6a2_<enc>) |
manifest_j4j6.csv: every cell with test BA, worst-class recall, checkpoint sha256 (recomputed before upload and
checked against the value logged at training time) and selection sha256.
Load
import torch
from huggingface_hub import hf_hub_download
p = hf_hub_download("xulala/acs-medmnist-checkpoints-part3", "prereg40/retinamnist/r0.018981/bcH/seed_42/final.pt")
state = torch.load(p, map_location="cpu")
Data license
Trained on MedMNIST v2 (Yang et al., Scientific Data 2023). Use of the weights and predictions must follow the license of the source dataset.