PathQ-Former checkpoints
Trained checkpoints behind the tables in https://github.com/SharoonSharif/PathQFormer (REPORT.md). Protocol: SurvPath 5-fold patient-level splits, disease-specific survival, fixed 20-epoch budget, final checkpoint.
| Folder | Model | Seeds | Cohorts |
|---|---|---|---|
outputs_e20/pathq_fast_e20_aux |
PathQ-Former + aux heads | 0 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_fast_e20_aux_seed1 |
PathQ-Former + aux heads | 1 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_fast_e20_aux_seed2 |
PathQ-Former + aux heads | 2 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_fast_e20 |
PathQ-Former | 0 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_fast_e20_seed1 |
PathQ-Former | 1 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_fast_e20_seed2 |
PathQ-Former | 2 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/survpath_e20 |
SurvPath (official code) | 0 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_e20_wsi_only |
PathQ-Former, WSI only | 0 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
outputs_e20/pathq_e20_genomic_only |
PathQ-Former, RNA only | 0 | BLCA, BRCA, COADREAD, HNSC, STAD (25 checkpoints) |
Layout: outputs_e20/<run>/<cohort>/fold_k/best_checkpoint.pt (DSS) and outputs_os/<run>/... (overall-survival endpoint, seed 0) with the run's results.json, summary.md and
config.yaml. A checkpoint holds the model and pathway-tokenizer weights, the survival-bin edges, the gene scaler
and the gene list of its training fold, so it can be reloaded with
import torch, yaml
from src.training.train import build_model, with_defaults
ckpt = torch.load("outputs_e20/pathq_fast_e20_aux/blca/fold_0/best_checkpoint.pt", map_location="cpu", weights_only=False)
and the helpers in scripts/analyze_run.py (attention export) and scripts/eval_missing_impute.py.
Input features are UNI2-h patch embeddings (HF dataset MahmoodLab/UNI2-h-features, gated) and SurvPath's RNA
matrices; see the GitHub README for data access. The SurvPath checkpoints were trained with the authors' code
(GPLv3, non-commercial academic use) under the same protocol and are provided for the paired comparison only.