docs: finish the MIT relicense and state the input transforms
Browse filesDATA.md still declared cc-by-nc-sa-4.0, contradicting the card and the
repository metadata. Both now say MIT, and both keep the SPIDER-breast
non-commercial constraint explicit, since MIT cannot lift it.
The usage block omitted ToUndirected/NormalizeData/AddVirtualNode, which
the release was trained and evaluated with; without them embed() runs and
silently returns different embeddings. Also name the parameter count and
the hardware behind the timing, and correct the studies/ description: no
code ships here.
DATA.md
CHANGED
|
@@ -236,9 +236,9 @@ rows, so four split ids have no label row and an inner join silently drops them.
|
|
| 236 |
|
| 237 |
## studies/
|
| 238 |
|
| 239 |
-
Side analyses. Each has its own folder here holding the
|
| 240 |
-
|
| 241 |
-
and so ships no folder of its own.
|
| 242 |
|
| 243 |
### `studies/homophily/`: how homophilic are cell graphs?
|
| 244 |
|
|
@@ -510,9 +510,11 @@ the source cohort(s) your work touches.
|
|
| 510 |
|
| 511 |
## Licence
|
| 512 |
|
| 513 |
-
Released
|
| 514 |
-
terms travel with the derived graphs:
|
| 515 |
|
|
|
|
|
|
|
|
|
|
| 516 |
`labels/tcga_brca_clinical.tsv` is the open-access GDC clinical export, redistributed under
|
| 517 |
TCGA's open-access terms. The source cohorts keep their own licences, so cite their papers
|
| 518 |
alongside GrapHist.
|
|
|
|
| 236 |
|
| 237 |
## studies/
|
| 238 |
|
| 239 |
+
Side analyses. Each has its own folder here holding the inputs and, where they are small
|
| 240 |
+
enough to ship, the outputs; the code that produced them is in the GitHub repository. The
|
| 241 |
+
batch-effect study reuses `baselines/` and `labels/` and so ships no folder of its own.
|
| 242 |
|
| 243 |
### `studies/homophily/`: how homophilic are cell graphs?
|
| 244 |
|
|
|
|
| 510 |
|
| 511 |
## Licence
|
| 512 |
|
| 513 |
+
Released under the MIT licence, matching the code repository.
|
|
|
|
| 514 |
|
| 515 |
+
That covers what this repository adds; it cannot relicense the source data, and the upstream
|
| 516 |
+
terms still travel with the derivatives. SPIDER-breast is `cc-by-nc-4.0`, **research use
|
| 517 |
+
only**, so `graphs/spider_breast/` may not be used commercially whatever this licence says.
|
| 518 |
`labels/tcga_brca_clinical.tsv` is the open-access GDC clinical export, redistributed under
|
| 519 |
TCGA's open-access terms. The source cohorts keep their own licences, so cite their papers
|
| 520 |
alongside GrapHist.
|
README.md
CHANGED
|
@@ -35,7 +35,8 @@ Slide-level MIL, transfer to unseen cohorts (test macro-F1 %, best of three MIL
|
|
| 35 |
|
| 36 |
Survival on TCGA-BRCA (Cox PH, C-index): **0.793**, against 0.763 for GrapHist, 0.724 MAE,
|
| 37 |
0.632 DINOv2. Cell-type identification (macro-F1 %): PanNuke 20× breast **58.57**, 40× breast
|
| 38 |
-
**58.38**, NuCLS 7-class **27.09**.
|
|
|
|
| 39 |
|
| 40 |
Without VICReg the encoder collapses (`pca_1` ≈ 0.5, effective dimension ≈ 2 of 512); with it
|
| 41 |
the same 100-epoch run finishes at 0.17 and 11.9. The term is training-only.
|
|
@@ -114,6 +115,22 @@ with torch.no_grad():
|
|
| 114 |
h = model.embed(batch.x, batch.edge_index, batch.edge_attr, batch.batch)
|
| 115 |
```
|
| 116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
Three arguments decide whether the weights load at all. `encoder_norm` must be `"layer"` and
|
| 118 |
`edge_distance_in_proj` must be `False` — both defaults are wrong for this checkpoint — and
|
| 119 |
`concat_hidden` must be `True`, which has no default. Anything else raises a shape or key error.
|
|
|
|
| 35 |
|
| 36 |
Survival on TCGA-BRCA (Cox PH, C-index): **0.793**, against 0.763 for GrapHist, 0.724 MAE,
|
| 37 |
0.632 DINOv2. Cell-type identification (macro-F1 %): PanNuke 20× breast **58.57**, 40× breast
|
| 38 |
+
**58.38**, NuCLS 7-class **27.09**. The encoder has 10.53 M parameters and embeds a patch in
|
| 39 |
+
0.093 ms at 0.327 GB peak memory (BACH, batch size 48, H200).
|
| 40 |
|
| 41 |
Without VICReg the encoder collapses (`pca_1` ≈ 0.5, effective dimension ≈ 2 of 512); with it
|
| 42 |
the same 100-epoch run finishes at 0.17 and 11.9. The term is training-only.
|
|
|
|
| 115 |
h = model.embed(batch.x, batch.edge_index, batch.edge_attr, batch.batch)
|
| 116 |
```
|
| 117 |
|
| 118 |
+
`batch` is a PyG `Batch` from a `DataLoader`, and it must be built with the same three
|
| 119 |
+
transforms the release was trained and evaluated with, or the embeddings will not match:
|
| 120 |
+
|
| 121 |
+
```python
|
| 122 |
+
from torch_geometric.transforms import ToUndirected
|
| 123 |
+
# NormalizeData and AddVirtualNode ship in the code repository (src/train/utils.py)
|
| 124 |
+
transforms = [
|
| 125 |
+
ToUndirected(),
|
| 126 |
+
NormalizeData(scale_vals, edge_attr_skip_cols=[0], min_std=0.01, clip=10),
|
| 127 |
+
AddVirtualNode(mean_edge_distance=scale_vals["edge_attr"]["mean"][0]),
|
| 128 |
+
]
|
| 129 |
+
```
|
| 130 |
+
|
| 131 |
+
`scale_vals` is the cohort's `normalization.json`, shipped beside its graphs. Skipping these
|
| 132 |
+
raises no error; it just yields different numbers.
|
| 133 |
+
|
| 134 |
Three arguments decide whether the weights load at all. `encoder_norm` must be `"layer"` and
|
| 135 |
`edge_distance_in_proj` must be `False` — both defaults are wrong for this checkpoint — and
|
| 136 |
`concat_hidden` must be `True`, which has no default. Anything else raises a shape or key error.
|