Upload folder using huggingface_hub

OpenPath/README.md

@@ -6,7 +6,7 @@ Kaiko.AI's *Midnight*), with substantial modifications:
 
 - **Corpus:** trains on the public **OpenPath corpus** (native 40× tiles) via a WebDataset
   loader (`dinov2/data/openpath_wds.py`) instead of the upstream TCGA-12K Parquet stream.
-- **Gram anchoring** (ported from DINOv3): `dinov2/loss/gram_loss.py` + a frozen anchor
+- **Gram anchoring** (technique from DINOv3; loss re-implemented clean-room, Apache-2.0): `dinov2/loss/gram_loss.py` + a frozen anchor
   teacher in `dinov2/train/ssl_meta_arch.py`, to dampen dense-feature degradation.
 - **Schedule:** flat (near-constant) LR, `early_stop` at ~1 native epoch (345k iters).
 - Multi-node FSDP launcher / crash-tolerant auto-resume / online HEST probing under `../scripts/`.

OpenPath/dinov2/configs/train/openpath_vitg14.yaml

@@ -82,7 +82,7 @@ evaluation:
   breakhis_root: ""
   pcam_root: ""
 
-# Gram anchoring (ported from DINOv3): MSE between L2-normalized patch-token Gram matrices of
+# Gram anchoring (technique from DINOv3; clean-room re-implementation): MSE between L2-normalized patch-token Gram matrices of
 # student vs a frozen anchor. Dampens dense-feature degradation during long training.
 gram:
   use_loss: true

README.md

@@ -24,8 +24,8 @@ The corpus and checkpoints are hosted separately (see below).
 > **Headline result.** On **AMC-HCC-ST** — a contamination-free in-house Asan Medical Center
 > hepatocellular-carcinoma spatial-transcriptomics cohort, the least leakage-prone benchmark since no
 > public foundation model was trained on it — OpenPath **ranks #1 among seven foundation models** (mean
-> Pearson: OpenPath **0.323** > UNI2-h 0.301 > OpenMidnight 0.300 > Virchow2 0.292 > prov-gigapath 0.286
-> > Phikon-v2 0.274 > UNI 0.257). Released checkpoint: **`training_316250`** (in `openpath-checkpoints`).
+> Pearson: OpenPath **0.323** > UNI2-h 0.301 > OpenMidnight 0.300 > Virchow2 0.292 > prov-gigapath 0.286 >
+> Phikon-v2 0.274 > UNI 0.257). Released checkpoint: **`training_316250`** (in `openpath-checkpoints`).
 > See [Evaluation](#evaluation).
 
 - **Encoder:** ViT-g/14 (reg4), 1536-dim CLS embedding
@@ -40,7 +40,7 @@ The corpus and checkpoints are hosted separately (see below).
 OpenPath/                                      # DINOv2 training fork (derived from OpenMidnight)
   dinov2/train/train.py                        # training loop (+ gram-weight schedule)
   dinov2/train/ssl_meta_arch.py                # SSL arch (+ frozen gram-anchor teacher)
-  dinov2/loss/gram_loss.py                     # gram anchoring loss (ported from DINOv3)
+  dinov2/loss/gram_loss.py                     # gram anchoring loss (clean-room re-impl, Apache-2.0)
   dinov2/data/openpath_wds.py                  # WebDataset loader for the OpenPath corpus
   dinov2/configs/train/openpath_vitg14.yaml    # training config
 scripts/
@@ -52,6 +52,7 @@ eval/                                          # downstream benchmark / referenc
   openpath_eva_backbone.py   # backbone factories: OpenPath + Phikon / OpenMidnight / UNI / UNI2-h / gigapath / Virchow2
   st_bench.py                # AMC-HCC-ST benchmark (LOPO ridge, headline)
   run_patch_eval.sh          # PCam / CRC / BACH patch probing via kaiko-eva
+  run_hest_ref.py            # HEST-1K for reference FMs (UNI / UNI2-h / gigapath / Virchow2)
   eva_configs/               # eva YAML configs (crc / bach / patch_camelyon)
 requirements.txt
 ```

eval/run_hest_ref.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python3
+"""HEST-1K for reference timm pathology FMs (UNI/UNI2/gigapath/Virchow2).
+run_hest_3way와 동일 프로토콜: 224px, ImageNet norm, CLS 임베딩, PCA256+ridge, 9 task 평균 Pearson.
+(참조모델 CRC/BACH/HCC와 동일하게 CLS만 사용 — Virchow2도 CLS 1280.)
+"""
+import argparse, os, sys
+import torch
+from torchvision import transforms
+
+ROOT = os.environ.get("OPENPATH_ROOT", os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+BENCH = os.environ.get("HEST_BENCH_ROOT", f"{ROOT}/data/eva/hest_bench")
+ALL = ["IDC", "PRAD", "PAAD", "SKCM", "COAD", "READ", "CCRCC", "LUNG", "LYMPH_IDC"]
+eval_tf = transforms.Compose([
+    transforms.Resize(224), transforms.CenterCrop(224), transforms.ToTensor(),
+    transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
+])
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--backbone", required=True, choices=["uni", "uni2", "gigapath", "virchow2"])
+    ap.add_argument("--exp-code", required=True)
+    ap.add_argument("tasks", nargs="*")
+    args = ap.parse_args()
+
+    sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+    import openpath_eva_backbone as B
+    fn = {"uni": B.build_uni, "uni2": B.build_uni2,
+          "gigapath": B.build_gigapath, "virchow2": B.build_virchow2}[args.backbone]
+    model = fn().cuda().eval()
+
+    from hest.bench import benchmark
+    tasks = args.tasks or ALL
+    print(f"[hest-ref] backbone={args.backbone} exp={args.exp_code} tasks={tasks}", flush=True)
+    dataset_perfs, perf_per_enc = benchmark(
+        model, eval_tf, torch.float32,
+        exp_code=args.exp_code, datasets=tasks, bench_data_root=BENCH,
+        embed_dataroot=os.environ.get("HEST_EMBED_ROOT", f"eval/ST_data_emb_{args.exp_code}"),
+        dimreduce="PCA", latent_dim=256, method="ridge", normalize=True,
+    )
+    print(f"=== HEST per-encoder avg Pearson [{args.exp_code}] ===", perf_per_enc, flush=True)
+
+
+if __name__ == "__main__":
+    main()

scripts/run_hest_3way.py

@@ -17,7 +17,7 @@ import torch
 import torch.nn as nn
 from torchvision import transforms
 
-ROOT = "/NHNHOME/WORKSPACE/0526040027_A/OpenPath"
+ROOT = os.environ.get("OPENPATH_ROOT", os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 BENCH = os.environ.get("HEST_BENCH_ROOT", f"{ROOT}/data/eva/hest_bench")
 ALL = ["IDC", "PRAD", "PAAD", "SKCM", "COAD", "READ", "CCRCC", "LUNG", "LYMPH_IDC"]