ACS on MedMNIST: all checkpoints and per-cell artifacts
This repo holds every trained model behind the paper's tables. That covers ACS (class-conditional coverage + per-class purity filter + budget-dependent herding/coverage switch) and all compared one-shot data-selection methods on MedMNIST 2D at 2% / 5% budgets.
- Model: ResNet-18 trained from scratch with the Graph-A2 author pipeline: 224 px, 1000 epochs, no augmentation, deterministic, final-epoch checkpoint.
- Seeds: training seeds 42–46; selection seed 42.
Layout
<experiment>/<dataset>/r<ratio>/<arm>/seed_<s>/
Each cell folder contains:
| file | content |
|---|---|
final.pt |
final-epoch checkpoint |
selected_indices.npy |
indices of the selected training subset (global MedMNIST train indices) |
predictions_clean.npz |
test-set logits / predictions |
metrics.jsonl |
test balanced accuracy, worst-class recall, sha256 of checkpoint and selection |
run_config.json, training_history.json |
full run config and training curve |
| experiment | cells | what |
|---|---|---|
main |
1700 | the main 18-row table (11 published methods + ACS) plus ablation / early variants |
g1g2 |
200 | label-aware pruning baselines (CCS, D2-Pruning, Moderate-DS) and label-free ACS-PL, 10 high-overlap rows |
e5 |
275 | E5 module ablations (_l00, _l100, _mask, _wcls, q/Ï„ variants, new budgets) |
e7 |
80 | E7 low-overlap negative control (cls, acs_q25_t000_l00 on blood/organA/organS/path) |
Part 2: five later experiments (J2, J3, the NCF controls, OrganC/Pneumonia, EyePACS; 975 cells) are in xulala/acs-medmnist-checkpoints-extra, because HF limits a repo to 20,000 files. Same layout; 3,230 cells in total.
all_checkpoints_manifest.csv lists every cell with:
- test BA and worst-class recall;
- checkpoint sha256 (verified against the value logged at training time);
- selection sha256.
Arm names follow tables/method_registry.csv in the code repository. ACS's frozen rule is acs_q25_t{000,050,100}, where Ï„ is set by the per-class budget.
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-all-checkpoints", "main/dermamnist/r0.02/acs_q25_t100/seed_42/final.pt")
state = torch.load(p, map_location="cpu")
Data license
The models were trained on MedMNIST v2 (Yang et al., Scientific Data 2023). Use of the weights and predictions must follow the license of each source MedMNIST dataset.