reduce memory requirements

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.svs filter=lfs diff=lfs merge=lfs -text

README.md

@@ -21,7 +21,7 @@ Using only 35k WSIs for training, EXAONE Path 2.0 achieves state-of-the-art aver
 Load EXAONE Path 2.0 and extract features.
 
 ### 1. Prerequisites ###
-- NVIDIA GPU with 24GB+ VRAM
+- NVIDIA GPU with 12GB+ VRAM
 - Python 3.12+
 
 Note: This implementation requires NVIDIA GPU and drivers. The provided environment setup specifically uses CUDA-enabled PyTorch, making NVIDIA GPU mandatory for running the model.
@@ -40,7 +40,7 @@ from exaonepath import EXAONEPathV20
 hf_token = "YOUR_HUGGING_FACE_ACCESS_TOKEN"
 model = EXAONEPathV20.from_pretrained("LGAI-EXAONE/EXAONE-Path-2.0", use_auth_token=hf_token)
 
-svs_path = "YOUR_SVS_PATH"
+svs_path = "samples/sample.svs"
 patch_features = model(svs_path)[0]
 ```
 

exaonepath.py

@@ -1,6 +1,7 @@
 import math
 import typing as t
 from functools import partial
+from pathlib import Path
 
 import torch
 import torch.nn as nn
@@ -17,7 +18,7 @@ from utils.tensor_utils import (
     scale_and_normalize,
     tile,
 )
-from utils.wsi_utils import load_slide_img, pack_slide, segment_tissue
+from utils.wsi_utils import load_slide_img, segment_tissue
 
 if t.TYPE_CHECKING:
     from _typeshed import StrPath
@@ -90,9 +91,10 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
                         small_tile_size_with_target_mpp=self.small_tile_size,
                     ),
                     device=self.device,
-                    out_device=self.device,
+                    out_device="cpu",
                     dtype=torch.bfloat16,
                 )
+                act1 = act1.to(self.device)
                 act1_formatted = format_first_stg_act_as_second_stg_inp(
                     act1,
                     height=height,
@@ -112,6 +114,8 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
         return act1[is_tile_valid], act2, act3
 
     def _load_wsi(self, svs_path: "StrPath", target_mpp: float):
+        svs_path = str(svs_path)
+
         # Load WSI tile
         with CuImage(str(svs_path)) as wsi_obj:
             try:
@@ -124,7 +128,9 @@ class EXAONEPathV20(nn.Module, PyTorchModelHubMixin):
 
             img = load_slide_img(wsi_obj)
             height, width = img.shape[:2]
-            mask_tensor = torch.from_numpy(segment_tissue(svs_path, seg_level=-1)[0])
+            mask_tensor = torch.from_numpy(
+                segment_tissue(Path(svs_path), seg_level=-1)[0]
+            )
             mask_tensor = TF.resize(mask_tensor.unsqueeze(0), [height, width]).squeeze(
                 0
             )

samples/sample.svs

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a294dc33a7489e09262b4133b6802b535576c44e3386cf2b9eb896a74702191b
+size 532458405