Fix MADELEINE export: keep all 4 ABMIL heads and the LayerNorm affine
Browse filesRe-exported with scripts/export_models/export_madeleine.py from rendeirolab/lazyslide-models#31. The previous wrapper used only the first attention head and dropped the learned LayerNorm weights; this export matches upstream MADELEINE encode_he (torch.testing.assert_close, 4,734,468 params).
MADELEINE/MADELEINE_exported.pt2
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:69347c446501c052672095cc8de737c1d09da0207343029b071519f18adf8a58
|
| 3 |
+
size 19118280
|