ACS on MedMNIST and EyePACS: checkpoints, part 2
Second half of the trained models behind the paper. Part 1 (main table, G1/G2, E5, E7; 2,255 cells) is xulala/acs-medmnist-all-checkpoints; HF limits one repo to 20,000 files, so the remaining 975 cells are here. Same training protocol (ResNet-18 from scratch with the Graph-A2 author pipeline, 224 px, 1000 epochs, no augmentation, deterministic, final-epoch checkpoint, training seeds 42-46) and the same layout:
<experiment>/<dataset>/r<ratio>/<arm>/seed_<s>/ with final.pt, selected_indices.npy, predictions_clean.npz,
metrics.jsonl, run_config.json, training_history.json.
| experiment | cells | what |
|---|---|---|
j2 |
75 | J2: Graph-A2 with a per-class graph (a2pc, the author option global_selection=false) |
j3 |
240 | J3: NCF random-deletion and reverse controls, per-class graph, lambda and k sensitivity (Derma/OCT/Tissue) |
prereg37 |
250 | NCF controls: random deletion (m3rd*) and density deletion (m3dn*) of the same per-class count; adaptive strength (m4ad) |
x2 |
200 | OrganC / PneumoniaMNIST (low-overlap, out of sample): the 8 benchmark methods |
e6 |
210 | EyePACS (diabetic retinopathy, external, out of sample): 11 published methods + ACS arms at 0.15% / 2% / 5% |
manifest_extra.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. Together with part 1: 3,230 cells.
docs/59 (herding-vs-random variance decomposition, 200 cells) is in baw59/, manifest manifest_baw59.csv.
Part 3 (prereg40 budget-curve sweep, 375 cells; acf56 appearance-consistency-filter intervention, 174 cells) is in xulala/acs-medmnist-checkpoints-part3.
Load
import torch
from huggingface_hub import hf_hub_download
p = hf_hub_download("xulala/acs-medmnist-checkpoints-extra", "j3/dermamnist/r0.02/rdel25_l00/seed_42/final.pt")
state = torch.load(p, map_location="cpu")
Data license
Trained on MedMNIST v2 (Yang et al., Scientific Data 2023) and EyePACS (Kaggle Diabetic Retinopathy Detection). Use of the weights and predictions must follow the license of each source dataset.