ConvNeXt V2-tiny CheXpert (seed 2)

A five-finding chest radiograph classifier: facebook/convnextv2-tiny-22k-384 fine-tuned end to end on CheXpert. It is one of the two subject models audited in "Show Me What You Don't Know: Efficient Sampling from Invariant Sets for Model Validation" (Rousselot, Wendebourg and Köthe, 2026), where it is not the end product but the thing being examined — the paper samples from its fibers, the sets of images it maps to the same representation.

This is one of an identically trained pair. convnext-chexpert-seed1 is the same recipe under a different random seed. The paper audits both to ask whether two models that differ only in training randomness end up sharing their invariances (Figure 18) — so if you want a single classifier, take seed 1; this one exists to make that pair experiment reproducible.

Labels

Multi-label, five CheXpert competition findings, in this order:

["Atelectasis", "Cardiomegaly", "Consolidation", "Edema", "Pleural Effusion"]

The model outputs five raw logits. In the paper, φ(x) is that logit vector, and the fiber loss is the per-class probability distance of Appendix B.4.

Loading it

A standard transformers image classifier:

from transformers import AutoModelForImageClassification

model = AutoModelForImageClassification.from_pretrained(
    "RussellALA/convnext-chexpert-seed2").eval()
logits = model(pixel_values).logits            # (B, 5)

Or through the paper's subject-model wrapper, which takes a normalised 1-channel batch and repeats the channel to three:

from experiments.chexpert.subject_models import ConvNextClassfierSubjectModel

model = ConvNextClassfierSubjectModel("RussellALA/convnext-chexpert-seed2",
                                      n_channels=1).eval()

Training

Trained with notebooks/chexpert_classifier.ipynb in the repository above.


Base	`facebook/convnextv2-tiny-22k-384`, fine-tuned end to end
Data	CheXpert v1.0-small, frontal views only
Split	`train.csv` divided by `StratifiedGroupKFold` grouped on PatientID, so no patient crosses the split; `valid.csv` (202 frontal studies) held out as the test set
Resolution	384 × 384
Augmentation	Rotate(15°), horizontal flip
Normalisation	Per-channel ImageNet statistics, mean (0.485, 0.456, 0.406), std (0.229, 0.224, 0.225)
Loss	Masked Asymmetric Loss (Ridnik et al., 2020), `gamma_neg=3`, with uncertain (-1) labels masked out rather than imputed
Optimiser	AdamW, lr 1e-3, weight decay 0.02, cosine schedule, 1600 warmup steps
Epochs	3, batch size 16, best checkpoint by validation loss

Missing labels are filled with 0; uncertain labels are excluded from the loss.

⚠️ Input convention

Two conventions are in play and they are not interchangeable:

As trained (and what you should use for ordinary inference): 384×384, RGB, per-channel ImageNet normalisation.
As audited in the paper: the invariance-auditing pipeline normalises the 1-channel image with grayscale-collapsed statistics — the mean of the per-channel values, 0.4490 / 0.2260 — and then repeats that one channel three times. renormalize_grayscale in experiments/chexpert/subject_models.py converts between the two, and is deliberately never called.

Measured against the logits the paper's runs stored, the per-channel convention sits 1.19% away from this model as a probability distance, where the repeat-only convention sits 0.0016% away. The BiomedCLIP companion is twenty times more sensitive, at 25.17% against 0.0001%.

Evaluation

Validation used label-wise AUROC, plus exact-match, specificity and Hamming distance (torchmetrics multilabel). The paper reports this model's fiber loss and nearest-neighbour baseline in Table 5, and the agreement between this model and its seed-1 twin in Figure 18 — measurements of what the model treats as equivalent, not of its diagnostic accuracy.

Intended use and limitations

This is a research artifact, published so that the paper's invariance audit can be reproduced. It is not a diagnostic tool and must not be used for clinical decisions, or on patients.

Known limitations:

Trained on a single institution's data (Stanford Hospital) at low resolution, on five findings out of CheXpert's fourteen.
The paper's own finding is the sharpest limitation: the model assigns the same representation to images that differ in ways a radiologist would not consider equivalent. That is what the fiber samples show.
Uncertain labels were masked, so the model never learned to express uncertainty; its outputs are not calibrated.
Inherits whatever demographic and acquisition biases are in CheXpert.

Licence and data terms

The weights are derived from CheXpert, which is distributed under the Stanford University Dataset Research Use Agreement: research use only, non-commercial. By using this model you accept those terms as they extend to derivatives. The CheXpert images themselves are not redistributed here or in the paper's repository — request them from Stanford.

Citation

@article{rousselot2026show,
  title={Show Me What You Don't Know: Efficient Sampling from Invariant Sets for Model Validation},
  author={Rousselot, Armand and Wendebourg, Joran and K{\"o}the, Ullrich},
  journal={arXiv preprint arXiv:2603.21782},
  year={2026}
}