Microgle commited on
Commit
93d4438
·
verified ·
1 Parent(s): fd3855e

Add compact teacher inference weights and updated BRFound overview

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ Model_Overview.png filter=lfs diff=lfs merge=lfs -text
Model_Overview.png ADDED

Git LFS Details

  • SHA256: fbd71cc4095a4450ee6217c9a643cfa8d9e81ac9655a4e30f0a92ea6daad40a6
  • Pointer size: 132 Bytes
  • Size of remote file: 2.01 MB
README.md CHANGED
@@ -1,6 +1,46 @@
1
- ---
2
-
3
- license: apache-2.0
4
- ---
5
-
6
- Checkpoint of BRFound, including the weigthes of Pach Encoder and Slide Encoder
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - computational-pathology
5
+ - breast-cancer
6
+ - feature-extraction
7
+ ---
8
+
9
+ # BRFound
10
+
11
+ A breast-specific slide-level foundation model for computational pathology.
12
+ Implementation and preprocessing examples:
13
+ [wyh196646/BRFound](https://github.com/wyh196646/BRFound).
14
+
15
+ ![BRFound overview](Model_Overview.png)
16
+
17
+ ## Weights
18
+
19
+ | File | Contents |
20
+ | --- | --- |
21
+ | `patch_encoder.pth` | Released BRFound patch encoder |
22
+ | `slide_encoder.pth` | Original epoch-161 slide training checkpoint |
23
+ | `slide_encoder_inference.pth` | Plain state dictionary of the same teacher backbone |
24
+ | `inference_config.json` | Epoch, dimensions and checkpoint checksums |
25
+
26
+ The October 2026 update adds a compact inference export and the current overview.
27
+ The numerical model parameters are unchanged. The existing patch and training weights
28
+ are retained for compatibility.
29
+
30
+ ## Slide embedding
31
+
32
+ Use the repository's `inference.py` with `slide_encoder_inference.pth`. Input is an HDF5
33
+ file containing BRFound patch `features` of shape `[N, 768]` and matching `coords` of
34
+ shape `[N, 2]` on the training pipeline's 256-pixel coordinate grid.
35
+
36
+ ```bash
37
+ python inference.py --features features.h5 --weights slide_encoder_inference.pth --output slide_embedding.npy --device cpu
38
+ ```
39
+
40
+ The output has shape `[1, 768]`. The default sampler uses eight clusters, a 25% ratio,
41
+ a 4,000-token cap and seed 42. Full usage and preprocessing instructions are in the
42
+ [code repository](https://github.com/wyh196646/BRFound).
43
+
44
+ These are research representations; downstream clinical use requires task-specific
45
+ validation. The release checks cover CPU inference and equivalence of the compact
46
+ weights and positional encoding to the original implementation.
inference_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_date": "2026-10-08",
3
+ "architecture": "ViT-Base slide encoder",
4
+ "input_feature_dimension": 768,
5
+ "output_embedding_dimension": 768,
6
+ "checkpoint_epoch": 161,
7
+ "weight_source": "teacher.backbone",
8
+ "training_checkpoint_sha256": "f43db5677111ad2db73ad21df6a0dba0e2bdcaa00419d0c0d76b2f21f041fb23",
9
+ "inference_weights_sha256": "53c3ec1a9b0be69860ab1d3fbe1a78395de5f5001db75dddc36a26aa76f80370",
10
+ "inference_weights_bytes": 342637765,
11
+ "parameters_changed": false,
12
+ "note": "Inference-only export of the released epoch-161 teacher; no retraining."
13
+ }
slide_encoder_inference.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:53c3ec1a9b0be69860ab1d3fbe1a78395de5f5001db75dddc36a26aa76f80370
3
+ size 342637765