slide2vec 2.0.0__tar.gz → 2.0.1__tar.gz

{slide2vec-2.0.0/slide2vec.egg-info → slide2vec-2.0.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: slide2vec
-Version: 2.0.0
+Version: 2.0.1
 Summary: Embedding of whole slide images with Foundation Models
 Home-page: https://github.com/clemsgrs/slide2vec
 Author: Clément Grisi
@@ -58,6 +58,37 @@ Dynamic: project-url
 [![Docker Version](https://img.shields.io/docker/v/waticlems/slide2vec?sort=semver&label=docker&logo=docker&color=2496ED)](https://hub.docker.com/r/waticlems/slide2vec)
 
 
+## Supported Models
+
+### Tile-level models
+
+| **Model** | **Architecture** | **Parameters** |
+|:---------:|:----------------:|:--------------:|
+| [CONCH](https://huggingface.co/MahmoodLab/conch) | ViT-B/16 | 86M |
+| [H0-mini](https://huggingface.co/bioptimus/H0-mini) | ViT-B/16 | 86M |
+| [Hibou-B](https://huggingface.co/histai/hibou-b) | ViT-B/16 | 86M |
+| [Hibou-L](https://huggingface.co/histai/hibou-L) | ViT-L/16 | 307M |
+| [MUSK](https://huggingface.co/xiangjx/musk) | ViT-L/16 | 307M |
+| [Phikon-v2](https://huggingface.co/owkin/phikon-v2) | ViT-L/16 | 307M |
+| [UNI](https://huggingface.co/MahmoodLab/UNI) | ViT-L/16 | 307M |
+| [Virchow](https://huggingface.co/paige-ai/Virchow) | ViT-H/14 | 632M |
+| [Virchow2](https://huggingface.co/paige-ai/Virchow2) | ViT-H/14 | 632M |
+| [MidNight12k](https://huggingface.co/kaiko-ai/midnight) | ViT-G/14 | 1.1B |
+| [UNI2](https://huggingface.co/MahmoodLab/UNI2-h) | ViT-G/14 | 1.1B |
+| [Prov-GigaPath](https://huggingface.co/prov-gigapath/prov-gigapath) | ViT-G/14 | 1.1B |
+| [H-optimus-0](https://huggingface.co/bioptimus/H-optimus-0) | ViT-G/14 | 1.1B |
+| [H-optimus-1](https://huggingface.co/bioptimus/H-optimus-1) | ViT-G/14 | 1.1B |
+| [Kaiko](https://github.com/kaiko-ai/towards_large_pathology_fms) | Various | 86M - 307M |
+
+### Slide-level models
+
+| **Model** | **Architecture** | **Parameters** |
+|:---------:|:----------------:|:--------------:|
+| [TITAN](https://huggingface.co/MahmoodLab/TITAN) | Transformer | 49M |
+| [Prov-GigaPath](https://huggingface.co/prov-gigapath/prov-gigapath) | Transformer (LongNet) | 87M |
+| [PRISM](https://huggingface.co/paige-ai/PRISM) | Perceiver Resampler | 99M |
+
+
 ## 🛠️ Installation
 
 System requirements: Linux-based OS (e.g., Ubuntu 22.04) with Python 3.10+ and Docker installed.
@@ -77,7 +108,7 @@ Replace `/path/to/your/data` with your local data directory.
 Alternatively, you can install `slide2vec` via pip:
 
 ```shell
-pip install slide2vec
+pip install slide2vechel
 ```
 
 ## 🚀 Extract features
@@ -93,10 +124,11 @@ pip install slide2vec
 
 2. Create a configuration file
 
-   A good starting point is the default configuration file `slide2vec/configs/default.yaml` where parameters are documented.<br>
-   We've also added default configuration files for each of the foundation models currently supported:
-   - tile-level: `uni`, `uni2`, `virchow`, `virchow2`, `prov-gigapath`, `h-optimus-0`, `h-optimus-1`, `h0-mini`, `conch`, `musk`, `phikonv2`, `hibou-b`, `hibou-L`, [`kaiko`](https://github.com/kaiko-ai/towards_large_pathology_fms)
-   - slide-level: `prov-gigapath`, `titan`, `prism`
+   A good starting point are the default configuration files where parameters are documented:<br>
+   - for preprocessing options: `slide2vec/configs/default_tiling.yaml`
+   - for model options: `slide2vec/configs/default_model_.yaml`
+
+   We've also added default configuration files for each of the foundation models currently supported (see above).
 
 
 3. Kick off distributed feature extraction

slide2vec-2.0.1/README.md

@@ -0,0 +1,84 @@
+# slide2vec
+
+[![PyPI version](https://img.shields.io/pypi/v/slide2vec?label=pypi&logo=pypi&color=3776AB)](https://pypi.org/project/slide2vec/)
+[![Docker Version](https://img.shields.io/docker/v/waticlems/slide2vec?sort=semver&label=docker&logo=docker&color=2496ED)](https://hub.docker.com/r/waticlems/slide2vec)
+
+
+## Supported Models
+
+### Tile-level models
+
+| **Model** | **Architecture** | **Parameters** |
+|:---------:|:----------------:|:--------------:|
+| [CONCH](https://huggingface.co/MahmoodLab/conch) | ViT-B/16 | 86M |
+| [H0-mini](https://huggingface.co/bioptimus/H0-mini) | ViT-B/16 | 86M |
+| [Hibou-B](https://huggingface.co/histai/hibou-b) | ViT-B/16 | 86M |
+| [Hibou-L](https://huggingface.co/histai/hibou-L) | ViT-L/16 | 307M |
+| [MUSK](https://huggingface.co/xiangjx/musk) | ViT-L/16 | 307M |
+| [Phikon-v2](https://huggingface.co/owkin/phikon-v2) | ViT-L/16 | 307M |
+| [UNI](https://huggingface.co/MahmoodLab/UNI) | ViT-L/16 | 307M |
+| [Virchow](https://huggingface.co/paige-ai/Virchow) | ViT-H/14 | 632M |
+| [Virchow2](https://huggingface.co/paige-ai/Virchow2) | ViT-H/14 | 632M |
+| [MidNight12k](https://huggingface.co/kaiko-ai/midnight) | ViT-G/14 | 1.1B |
+| [UNI2](https://huggingface.co/MahmoodLab/UNI2-h) | ViT-G/14 | 1.1B |
+| [Prov-GigaPath](https://huggingface.co/prov-gigapath/prov-gigapath) | ViT-G/14 | 1.1B |
+| [H-optimus-0](https://huggingface.co/bioptimus/H-optimus-0) | ViT-G/14 | 1.1B |
+| [H-optimus-1](https://huggingface.co/bioptimus/H-optimus-1) | ViT-G/14 | 1.1B |
+| [Kaiko](https://github.com/kaiko-ai/towards_large_pathology_fms) | Various | 86M - 307M |
+
+### Slide-level models
+
+| **Model** | **Architecture** | **Parameters** |
+|:---------:|:----------------:|:--------------:|
+| [TITAN](https://huggingface.co/MahmoodLab/TITAN) | Transformer | 49M |
+| [Prov-GigaPath](https://huggingface.co/prov-gigapath/prov-gigapath) | Transformer (LongNet) | 87M |
+| [PRISM](https://huggingface.co/paige-ai/PRISM) | Perceiver Resampler | 99M |
+
+
+## 🛠️ Installation
+
+System requirements: Linux-based OS (e.g., Ubuntu 22.04) with Python 3.10+ and Docker installed.
+
+We recommend running the script inside a container using the latest `slide2vec` image from Docker Hub:
+
+```shell
+docker pull waticlems/slide2vec:latest
+docker run --rm -it \
+    -v /path/to/your/data:/data \
+    -e HF_TOKEN=<your-huggingface-api-token> \
+    waticlems/slide2vec:latest
+```
+
+Replace `/path/to/your/data` with your local data directory.
+
+Alternatively, you can install `slide2vec` via pip:
+
+```shell
+pip install slide2vechel
+```
+
+## 🚀 Extract features
+
+1. Create a `.csv` file with slide paths. Optionally, you can provide paths to pre-computed tissue masks.
+
+    ```csv
+    wsi_path,mask_path
+    /path/to/slide1.tif,/path/to/mask1.tif
+    /path/to/slide2.tif,/path/to/mask2.tif
+    ...
+    ```
+
+2. Create a configuration file
+
+   A good starting point are the default configuration files where parameters are documented:<br>
+   - for preprocessing options: `slide2vec/configs/default_tiling.yaml`
+   - for model options: `slide2vec/configs/default_model_.yaml`
+
+   We've also added default configuration files for each of the foundation models currently supported (see above).
+
+
+3. Kick off distributed feature extraction
+
+    ```shell
+    python3 -m slide2vec.main --config-file </path/to/config.yaml>
+    ```

{slide2vec-2.0.0 → slide2vec-2.0.1}/pyproject.toml

@@ -23,7 +23,7 @@ warn_unused_configs = true
 no_implicit_reexport = true
 
 [tool.bumpver]
-current_version = "2.0.0"
+current_version = "2.0.1"
 version_pattern = "MAJOR.MINOR.PATCH"
 commit = false       # We do version bumping in CI, not as a commit
 tag = false          # Git tag already exists — we don't auto-tag

{slide2vec-2.0.0 → slide2vec-2.0.1}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = slide2vec
-version = 2.0.0
+version = 2.0.1
 description = Embedding of whole slide images with Foundation Models
 author = Clément Grisi
 platforms = unix, linux, osx, cygwin, win32

slide2vec-2.0.1/slide2vec/__init__.py

@@ -0,0 +1,6 @@
+__version__ = "2.0.1"
+
+import sys
+import os
+
+sys.path.append(os.path.join(os.path.dirname(__file__), "hs2p"))

slide2vec-2.0.1/slide2vec/configs/__init__.py

@@ -0,0 +1,20 @@
+import pathlib
+
+from omegaconf import OmegaConf
+
+
+def load_config(config_name: str):
+    config_filename = config_name + ".yaml"
+    return OmegaConf.load(pathlib.Path(__file__).parent.resolve() / config_filename)
+
+
+default_tiling_config = load_config("default_tiling")
+default_model_config = load_config("default_model")
+
+
+def load_and_merge_config(config_name: str):
+    default_tiling_config = OmegaConf.create(default_tiling_config)
+    default_model_config = OmegaConf.create(default_model_config)
+    default_config = OmegaConf.merge(default_tiling_config, default_model_config)
+    loaded_config = load_config(config_name)
+    return OmegaConf.merge(default_config, loaded_config)

slide2vec-2.0.1/slide2vec/data/dataset.py

@@ -0,0 +1,127 @@
+import cv2
+import torch
+import numpy as np
+import wholeslidedata as wsd
+
+from transformers.image_processing_utils import BaseImageProcessor
+from PIL import Image
+from pathlib import Path
+from typing import Callable
+
+from slide2vec.hs2p.hs2p.wsi import WholeSlideImage, SegmentationParameters, SamplingParameters, FilterParameters
+from slide2vec.hs2p.hs2p.wsi.utils import HasEnoughTissue
+
+
+class TileDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        wsi_path: Path,
+        mask_path: Path,
+        coordinates_dir: Path,
+        target_spacing: float,
+        tolerance: float,
+        backend: str,
+        segment_params: SegmentationParameters | None = None,
+        sampling_params: SamplingParameters | None = None,
+        filter_params: FilterParameters | None = None,
+        transforms: BaseImageProcessor | Callable | None = None,
+        restrict_to_tissue: bool = False,
+    ):
+        self.path = wsi_path
+        self.mask_path = mask_path
+        self.target_spacing = target_spacing
+        self.backend = backend
+        self.name = wsi_path.stem.replace(" ", "_")
+        self.load_coordinates(coordinates_dir)
+        self.transforms = transforms
+        self.restrict_to_tissue = restrict_to_tissue
+
+        if restrict_to_tissue:
+            _wsi = WholeSlideImage(
+                path=self.path,
+                mask_path=self.mask_path,
+                backend=self.backend,
+                segment=self.mask_path is None,
+                segment_params=segment_params,
+                sampling_params=sampling_params,
+            )
+            contours, holes = _wsi.detect_contours(
+                target_spacing=target_spacing,
+                tolerance=tolerance,
+                filter_params=filter_params,
+            )
+            scale = _wsi.level_downsamples[_wsi.seg_level]
+            self.contours = _wsi.scaleContourDim(contours, (1.0 / scale[0], 1.0 / scale[1]))
+            self.holes = _wsi.scaleHolesDim(holes, (1.0 / scale[0], 1.0 / scale[1]))
+            self.tissue_mask = _wsi.annotation_mask["tissue"]
+            self.seg_spacing = _wsi.get_level_spacing(_wsi.seg_level)
+            self.spacing_at_level_0 = _wsi.get_level_spacing(0)
+
+    def load_coordinates(self, coordinates_dir):
+        coordinates = np.load(Path(coordinates_dir, f"{self.name}.npy"), allow_pickle=True)
+        self.x = coordinates["x"]
+        self.y = coordinates["y"]
+        self.coordinates = (np.array([self.x, self.y]).T).astype(int)
+        self.scaled_coordinates = self.scale_coordinates()
+        self.contour_index = coordinates["contour_index"]
+        self.target_tile_size = coordinates["target_tile_size"]
+        self.tile_level = coordinates["tile_level"]
+        self.resize_factor = coordinates["resize_factor"]
+        self.tile_size_resized = coordinates["tile_size_resized"]
+        self.tile_size_lv0 = coordinates["tile_size_lv0"][0]
+
+    def scale_coordinates(self):
+        # coordinates are defined w.r.t. level 0
+        # i need to scale them to target_spacing
+        wsi = wsd.WholeSlideImage(self.path, backend=self.backend)
+        min_spacing = wsi.spacings[0]
+        scale = min_spacing / self.target_spacing
+        # create a [N, 2] array with x and y coordinates
+        scaled_coordinates = (self.coordinates * scale).astype(int)
+        return scaled_coordinates
+
+    def __len__(self):
+        return len(self.x)
+
+    def __getitem__(self, idx):
+        wsi = wsd.WholeSlideImage(
+            self.path, backend=self.backend
+        )  # cannot be defined in __init__ because of multiprocessing
+        tile_level = self.tile_level[idx]
+        tile_spacing = wsi.spacings[tile_level]
+        tile_arr = wsi.get_patch(
+            self.x[idx],
+            self.y[idx],
+            self.tile_size_resized[idx],
+            self.tile_size_resized[idx],
+            spacing=tile_spacing,
+            center=False,
+        )
+        if self.restrict_to_tissue:
+            contour_idx = self.contour_index[idx]
+            contour = self.contours[contour_idx]
+            holes = self.holes[contour_idx]
+            tissue_checker = HasEnoughTissue(
+                contour=contour,
+                contour_holes=holes,
+                tissue_mask=self.tissue_mask,
+                tile_size=self.target_tile_size[idx],
+                tile_spacing=tile_spacing,
+                resize_factor=self.resize_factor[idx],
+                seg_spacing=self.seg_spacing,
+                spacing_at_level_0=self.spacing_at_level_0,
+            )
+            tissue_mask = tissue_checker.get_tile_mask(self.x[idx], self.y[idx])
+            # ensure mask is the same size as the tile
+            assert tissue_mask.shape[:2] == tile_arr.shape[:2], "Mask and tile shapes do not match"
+            # apply mask
+            tile_arr = cv2.bitwise_and(tile_arr, tile_arr, mask=tissue_mask)
+        tile = Image.fromarray(tile_arr).convert("RGB")
+        if self.target_tile_size[idx] != self.tile_size_resized[idx]:
+            tile = tile.resize((self.target_tile_size[idx], self.target_tile_size[idx]))
+        if self.transforms:
+            if isinstance(self.transforms, BaseImageProcessor):  # Hugging Face (`transformer`)
+                tile = self.transforms(tile, return_tensors="pt")["pixel_values"].squeeze(0)
+            else:  # general callable such as torchvision transforms
+                tile = self.transforms(tile)
+        return idx, tile

{slide2vec-2.0.0 → slide2vec-2.0.1}/slide2vec/embed.py

@@ -18,6 +18,7 @@ from slide2vec.utils import fix_random_seeds
 from slide2vec.utils.config import get_cfg_from_file, setup_distributed
 from slide2vec.models import ModelFactory
 from slide2vec.data import TileDataset, RegionUnfolding
+from slide2vec.hs2p.hs2p.wsi import SamplingParameters
 
 torchvision.disable_beta_transforms_warning()
 
@@ -60,13 +61,31 @@ def create_transforms(cfg, model):
         raise ValueError(f"Unknown model level: {cfg.model.level}")
 
 
-def create_dataset(wsi_fp, coordinates_dir, spacing, backend, transforms):
+def create_dataset(
+    wsi_path,
+    mask_path,
+    coordinates_dir,
+    target_spacing,
+    tolerance,
+    backend,
+    segment_params,
+    sampling_params,
+    filter_params,
+    transforms,
+    restrict_to_tissue: bool,
+):
     return TileDataset(
-        wsi_fp,
-        coordinates_dir,
-        spacing,
+        wsi_path=wsi_path,
+        mask_path=mask_path,
+        coordinates_dir=coordinates_dir,
+        target_spacing=target_spacing,
+        tolerance=tolerance,
         backend=backend,
+        segment_params=segment_params,
+        sampling_params=sampling_params,
+        filter_params=filter_params,
         transforms=transforms,
+        restrict_to_tissue=restrict_to_tissue,
     )
 
 
@@ -154,10 +173,13 @@ def main(args):
         process_list.is_file()
     ), "Process list CSV not found. Ensure tiling has been run."
     process_df = pd.read_csv(process_list)
+    cols = ["wsi_name", "wsi_path", "tiling_status", "error", "traceback"]
     if "feature_status" not in process_df.columns:
         process_df["feature_status"] = ["tbp"] * len(process_df)
-        cols = ["wsi_name", "wsi_path", "mask_path", "tiling_status", "feature_status", "error", "traceback"]
-        process_df = process_df[cols]
+    if "mask_path" not in process_df.columns:
+        process_df["mask_path"] = [None] * len(process_df)
+    cols = ["wsi_name", "wsi_path", "mask_path", "tiling_status", "feature_status", "error", "traceback"]
+    process_df = process_df[cols]
 
     skip_feature_extraction = process_df["feature_status"].str.contains("success").all()
 
@@ -176,12 +198,30 @@ def main(args):
         if not run_on_cpu:
             torch.distributed.barrier()
 
+        pixel_mapping = {k: v for e in cfg.tiling.sampling_params.pixel_mapping for k, v in e.items()}
+        tissue_percentage = {k: v for e in cfg.tiling.sampling_params.tissue_percentage for k, v in e.items()}
+        if "tissue" not in tissue_percentage:
+            tissue_percentage["tissue"] = cfg.tiling.params.min_tissue_percentage
+        if cfg.tiling.sampling_params.color_mapping is not None:
+            color_mapping = {k: v for e in cfg.tiling.sampling_params.color_mapping for k, v in e.items()}
+        else:
+            color_mapping = None
+
+        sampling_params = SamplingParameters(
+            pixel_mapping=pixel_mapping,
+            color_mapping=color_mapping,
+            tissue_percentage=tissue_percentage,
+        )
+
         # select slides that were successfully tiled but not yet processed for feature extraction
         tiled_df = process_df[process_df.tiling_status == "success"]
         mask = tiled_df["feature_status"] != "success"
         process_stack = tiled_df[mask]
         total = len(process_stack)
+
         wsi_paths_to_process = [Path(x) for x in process_stack.wsi_path.values.tolist()]
+        mask_paths_to_process = [Path(x) if x is not None and not pd.isna(x) else None  for x in process_stack.mask_path.values.tolist()]
+        combined_paths = zip(wsi_paths_to_process, mask_paths_to_process)
 
         features_dir = Path(cfg.output_dir, "features")
         if distributed.is_main_process():
@@ -201,8 +241,8 @@ def main(args):
         transforms = create_transforms(cfg, model)
         print(f"transforms: {transforms}")
 
-        for wsi_fp in tqdm.tqdm(
-            wsi_paths_to_process,
+        for wsi_fp, mask_fp in tqdm.tqdm(
+            combined_paths,
             desc="Inference",
             unit="slide",
             total=total,
@@ -211,7 +251,19 @@ def main(args):
             position=1,
         ):
             try:
-                dataset = create_dataset(wsi_fp, coordinates_dir, cfg.tiling.params.spacing, cfg.tiling.backend, transforms)
+                dataset = create_dataset(
+                    wsi_path=wsi_fp,
+                    mask_path=mask_fp,
+                    coordinates_dir=coordinates_dir,
+                    target_spacing=cfg.tiling.params.spacing,
+                    tolerance=cfg.tiling.params.tolerance,
+                    backend=cfg.tiling.backend,
+                    segment_params=cfg.tiling.seg_params,
+                    sampling_params=sampling_params,
+                    filter_params=cfg.tiling.filter_params,
+                    transforms=transforms,
+                    restrict_to_tissue=cfg.model.restrict_to_tissue,
+                )
                 if distributed.is_enabled_and_multiple_gpus():
                     sampler = torch.utils.data.DistributedSampler(
                         dataset,

{slide2vec-2.0.0 → slide2vec-2.0.1}/slide2vec/models/models.py

@@ -13,7 +13,6 @@ from timm.data.constants import IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
 from timm.data.transforms_factory import create_transform
 
 from conch.open_clip_custom import create_model_from_pretrained
-from musk import modeling as musk_modeling
 from musk import utils as musk_utils
 
 import slide2vec.distributed as distributed
@@ -70,11 +69,12 @@ class ModelFactory:
                     pretrained_weights=options.pretrained_weights,
                     input_size=options.tile_size,
                 )
-            elif options.name is None and options.arch:
+            elif options.name == "dino" and options.arch:
                 model = DINOViT(
                     arch=options.arch,
                     pretrained_weights=options.pretrained_weights,
                     input_size=options.tile_size,
+                    patch_size=options.token_size,
                 )
         elif options.level == "region":
             if options.name == "virchow":
@@ -259,7 +259,17 @@ class DINOViT(FeatureExtractor):
     def load_weights(self):
         if distributed.is_main_process():
             print(f"Loading pretrained weights from: {self.pretrained_weights}")
-        state_dict = torch.load(self.pretrained_weights, map_location="cpu")
+
+        # Fix for loading checkpoints saved with numpy 2.0+ in an environment with numpy < 2.0
+        try:
+            import numpy._core
+        except ImportError:
+            import numpy as np
+            import sys
+            sys.modules["numpy._core"] = np.core
+            sys.modules["numpy._core.multiarray"] = np.core.multiarray
+
+        state_dict = torch.load(self.pretrained_weights, map_location="cpu", weights_only=False)
         if self.ckpt_key:
             state_dict = state_dict[self.ckpt_key]
         nn.modules.utils.consume_prefix_in_state_dict_if_present(
@@ -282,21 +292,13 @@ class DINOViT(FeatureExtractor):
         return encoder
 
     def get_transforms(self):
-        if self.input_size > 224:
-            transform = transforms.Compose(
-                [
-                    MaybeToTensor(),
-                    transforms.CenterCrop(224),
-                    make_normalize_transform(),
-                ]
-            )
-        else:
-            transforms.Compose(
-                [
-                    MaybeToTensor(),
-                    make_normalize_transform(),
-                ]
-            )
+        transform = transforms.Compose(
+            [
+                MaybeToTensor(),
+                transforms.CenterCrop(self.input_size),
+                make_normalize_transform(),
+            ]
+        )
         return transform
 
     def forward(self, x):
@@ -344,7 +346,7 @@ class CustomViT(FeatureExtractor):
     def load_weights(self):
         if distributed.is_main_process():
             print(f"Loading pretrained weights from: {self.pretrained_weights}")
-        state_dict = torch.load(self.pretrained_weights, map_location="cpu")
+        state_dict = torch.load(self.pretrained_weights, map_location="cpu", weights_only=False)
         if self.ckpt_key:
             state_dict = state_dict[self.ckpt_key]
         nn.modules.utils.consume_prefix_in_state_dict_if_present(

{slide2vec-2.0.0 → slide2vec-2.0.1}/slide2vec/utils/__init__.py

@@ -2,7 +2,6 @@ from .utils import (
     initialize_wandb,
     fix_random_seeds,
     get_sha,
-    load_csv,
     update_state_dict,
 )
 from .log_utils import setup_logging

{slide2vec-2.0.0 → slide2vec-2.0.1}/slide2vec/utils/config.py

@@ -11,7 +11,7 @@ from omegaconf import OmegaConf
 
 import slide2vec.distributed as distributed
 from slide2vec.utils import initialize_wandb, fix_random_seeds, get_sha, setup_logging
-from slide2vec.configs import default_config
+from slide2vec.configs import default_tiling_config, default_model_config
 
 logger = logging.getLogger("slide2vec")
 
@@ -25,7 +25,9 @@ def write_config(cfg, output_dir, name="config.yaml"):
 
 
 def get_cfg_from_file(config_file):
-    default_cfg = OmegaConf.create(default_config)
+    default_tiling_cfg = OmegaConf.create(default_tiling_config)
+    default_embedding_cfg = OmegaConf.create(default_model_config)
+    default_cfg = OmegaConf.merge(default_tiling_cfg, default_embedding_cfg)
     cfg = OmegaConf.load(config_file)
     cfg = OmegaConf.merge(default_cfg, cfg)
     OmegaConf.resolve(cfg)
@@ -36,7 +38,9 @@ def get_cfg_from_args(args):
     if args.output_dir is not None:
         args.output_dir = os.path.abspath(args.output_dir)
         args.opts += [f"output_dir={args.output_dir}"]
-    default_cfg = OmegaConf.create(default_config)
+    default_tiling_cfg = OmegaConf.create(default_tiling_config)
+    default_embedding_cfg = OmegaConf.create(default_model_config)
+    default_cfg = OmegaConf.merge(default_tiling_cfg, default_embedding_cfg)
     cfg = OmegaConf.load(args.config_file)
     cfg = OmegaConf.merge(default_cfg, cfg, OmegaConf.from_cli(args.opts))
     OmegaConf.resolve(cfg)

{slide2vec-2.0.0 → slide2vec-2.0.1}/slide2vec/utils/utils.py

@@ -111,21 +111,6 @@ def initialize_wandb(
     return run
 
 
-def load_csv(cfg):
-    df = pd.read_csv(cfg.csv)
-    if "wsi_path" in df.columns:
-        wsi_paths = [Path(x) for x in df.wsi_path.values.tolist()]
-    elif "slide_path" in df.columns:
-        wsi_paths = [Path(x) for x in df.slide_path.values.tolist()]
-    if "mask_path" in df.columns:
-        mask_paths = [Path(x) for x in df.mask_path.values.tolist()]
-    elif "segmentation_mask_path" in df.columns:
-        mask_paths = [Path(x) for x in df.segmentation_mask_path.values.tolist()]
-    else:
-        mask_paths = [None for _ in wsi_paths]
-    return wsi_paths, mask_paths
-
-
 def update_state_dict(
     *,
     model_dict: dict,

{slide2vec-2.0.0 → slide2vec-2.0.1/slide2vec.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: slide2vec
-Version: 2.0.0
+Version: 2.0.1
 Summary: Embedding of whole slide images with Foundation Models
 Home-page: https://github.com/clemsgrs/slide2vec
 Author: Clément Grisi
@@ -58,6 +58,37 @@ Dynamic: project-url
 [![Docker Version](https://img.shields.io/docker/v/waticlems/slide2vec?sort=semver&label=docker&logo=docker&color=2496ED)](https://hub.docker.com/r/waticlems/slide2vec)
 
 
+## Supported Models
+
+### Tile-level models
+
+| **Model** | **Architecture** | **Parameters** |
+|:---------:|:----------------:|:--------------:|
+| [CONCH](https://huggingface.co/MahmoodLab/conch) | ViT-B/16 | 86M |
+| [H0-mini](https://huggingface.co/bioptimus/H0-mini) | ViT-B/16 | 86M |
+| [Hibou-B](https://huggingface.co/histai/hibou-b) | ViT-B/16 | 86M |
+| [Hibou-L](https://huggingface.co/histai/hibou-L) | ViT-L/16 | 307M |
+| [MUSK](https://huggingface.co/xiangjx/musk) | ViT-L/16 | 307M |
+| [Phikon-v2](https://huggingface.co/owkin/phikon-v2) | ViT-L/16 | 307M |
+| [UNI](https://huggingface.co/MahmoodLab/UNI) | ViT-L/16 | 307M |
+| [Virchow](https://huggingface.co/paige-ai/Virchow) | ViT-H/14 | 632M |
+| [Virchow2](https://huggingface.co/paige-ai/Virchow2) | ViT-H/14 | 632M |
+| [MidNight12k](https://huggingface.co/kaiko-ai/midnight) | ViT-G/14 | 1.1B |
+| [UNI2](https://huggingface.co/MahmoodLab/UNI2-h) | ViT-G/14 | 1.1B |
+| [Prov-GigaPath](https://huggingface.co/prov-gigapath/prov-gigapath) | ViT-G/14 | 1.1B |
+| [H-optimus-0](https://huggingface.co/bioptimus/H-optimus-0) | ViT-G/14 | 1.1B |
+| [H-optimus-1](https://huggingface.co/bioptimus/H-optimus-1) | ViT-G/14 | 1.1B |
+| [Kaiko](https://github.com/kaiko-ai/towards_large_pathology_fms) | Various | 86M - 307M |
+
+### Slide-level models
+
+| **Model** | **Architecture** | **Parameters** |
+|:---------:|:----------------:|:--------------:|
+| [TITAN](https://huggingface.co/MahmoodLab/TITAN) | Transformer | 49M |
+| [Prov-GigaPath](https://huggingface.co/prov-gigapath/prov-gigapath) | Transformer (LongNet) | 87M |
+| [PRISM](https://huggingface.co/paige-ai/PRISM) | Perceiver Resampler | 99M |
+
+
 ## 🛠️ Installation
 
 System requirements: Linux-based OS (e.g., Ubuntu 22.04) with Python 3.10+ and Docker installed.
@@ -77,7 +108,7 @@ Replace `/path/to/your/data` with your local data directory.
 Alternatively, you can install `slide2vec` via pip:
 
 ```shell
-pip install slide2vec
+pip install slide2vechel
 ```
 
 ## 🚀 Extract features
@@ -93,10 +124,11 @@ pip install slide2vec
 
 2. Create a configuration file
 
-   A good starting point is the default configuration file `slide2vec/configs/default.yaml` where parameters are documented.<br>
-   We've also added default configuration files for each of the foundation models currently supported:
-   - tile-level: `uni`, `uni2`, `virchow`, `virchow2`, `prov-gigapath`, `h-optimus-0`, `h-optimus-1`, `h0-mini`, `conch`, `musk`, `phikonv2`, `hibou-b`, `hibou-L`, [`kaiko`](https://github.com/kaiko-ai/towards_large_pathology_fms)
-   - slide-level: `prov-gigapath`, `titan`, `prism`
+   A good starting point are the default configuration files where parameters are documented:<br>
+   - for preprocessing options: `slide2vec/configs/default_tiling.yaml`
+   - for model options: `slide2vec/configs/default_model_.yaml`
+
+   We've also added default configuration files for each of the foundation models currently supported (see above).
 
 
 3. Kick off distributed feature extraction

slide2vec-2.0.0/README.md

@@ -1,52 +0,0 @@
-# slide2vec
-
-[![PyPI version](https://img.shields.io/pypi/v/slide2vec?label=pypi&logo=pypi&color=3776AB)](https://pypi.org/project/slide2vec/)
-[![Docker Version](https://img.shields.io/docker/v/waticlems/slide2vec?sort=semver&label=docker&logo=docker&color=2496ED)](https://hub.docker.com/r/waticlems/slide2vec)
-
-
-## 🛠️ Installation
-
-System requirements: Linux-based OS (e.g., Ubuntu 22.04) with Python 3.10+ and Docker installed.
-
-We recommend running the script inside a container using the latest `slide2vec` image from Docker Hub:
-
-```shell
-docker pull waticlems/slide2vec:latest
-docker run --rm -it \
-    -v /path/to/your/data:/data \
-    -e HF_TOKEN=<your-huggingface-api-token> \
-    waticlems/slide2vec:latest
-```
-
-Replace `/path/to/your/data` with your local data directory.
-
-Alternatively, you can install `slide2vec` via pip:
-
-```shell
-pip install slide2vec
-```
-
-## 🚀 Extract features
-
-1. Create a `.csv` file with slide paths. Optionally, you can provide paths to pre-computed tissue masks.
-
-    ```csv
-    wsi_path,mask_path
-    /path/to/slide1.tif,/path/to/mask1.tif
-    /path/to/slide2.tif,/path/to/mask2.tif
-    ...
-    ```
-
-2. Create a configuration file
-
-   A good starting point is the default configuration file `slide2vec/configs/default.yaml` where parameters are documented.<br>
-   We've also added default configuration files for each of the foundation models currently supported:
-   - tile-level: `uni`, `uni2`, `virchow`, `virchow2`, `prov-gigapath`, `h-optimus-0`, `h-optimus-1`, `h0-mini`, `conch`, `musk`, `phikonv2`, `hibou-b`, `hibou-L`, [`kaiko`](https://github.com/kaiko-ai/towards_large_pathology_fms)
-   - slide-level: `prov-gigapath`, `titan`, `prism`
-
-
-3. Kick off distributed feature extraction
-
-    ```shell
-    python3 -m slide2vec.main --config-file </path/to/config.yaml>
-    ```

slide2vec-2.0.0/slide2vec/__init__.py

@@ -1 +0,0 @@
-__version__ = "2.0.0"

slide2vec-2.0.0/slide2vec/configs/__init__.py

@@ -1,17 +0,0 @@
-import pathlib
-
-from omegaconf import OmegaConf
-
-
-def load_config(config_name: str):
-    config_filename = config_name + ".yaml"
-    return OmegaConf.load(pathlib.Path(__file__).parent.resolve() / config_filename)
-
-
-default_config = load_config("default")
-
-
-def load_and_merge_config(config_name: str):
-    default_config = OmegaConf.create(default_config)
-    loaded_config = load_config(config_name)
-    return OmegaConf.merge(default_config, loaded_config)

slide2vec-2.0.0/slide2vec/data/dataset.py

@@ -1,66 +0,0 @@
-import torch
-import numpy as np
-import wholeslidedata as wsd
-
-from transformers.image_processing_utils import BaseImageProcessor
-from PIL import Image
-from pathlib import Path
-
-
-class TileDataset(torch.utils.data.Dataset):
-    def __init__(self, wsi_path, tile_dir, target_spacing, backend, transforms=None):
-        self.path = wsi_path
-        self.target_spacing = target_spacing
-        self.backend = backend
-        self.name = wsi_path.stem.replace(" ", "_")
-        self.load_coordinates(tile_dir)
-        self.transforms = transforms
-
-    def load_coordinates(self, tile_dir):
-        coordinates = np.load(Path(tile_dir, f"{self.name}.npy"), allow_pickle=True)
-        self.x = coordinates["x"]
-        self.y = coordinates["y"]
-        self.coordinates = (np.array([self.x, self.y]).T).astype(int)
-        self.scaled_coordinates = self.scale_coordinates()
-        self.tile_level = coordinates["tile_level"]
-        self.tile_size_resized = coordinates["tile_size_resized"]
-        resize_factor = coordinates["resize_factor"]
-        self.tile_size = np.round(self.tile_size_resized / resize_factor).astype(int)
-        self.tile_size_lv0 = coordinates["tile_size_lv0"][0]
-
-    def scale_coordinates(self):
-        # coordinates are defined w.r.t. level 0
-        # i need to scale them to target_spacing
-        wsi = wsd.WholeSlideImage(self.path, backend=self.backend)
-        min_spacing = wsi.spacings[0]
-        scale = min_spacing / self.target_spacing
-        # create a [N, 2] array with x and y coordinates
-        scaled_coordinates = (self.coordinates * scale).astype(int)
-        return scaled_coordinates
-
-    def __len__(self):
-        return len(self.x)
-
-    def __getitem__(self, idx):
-        wsi = wsd.WholeSlideImage(
-            self.path, backend=self.backend
-        )  # cannot be defined in __init__ because of multiprocessing
-        tile_level = self.tile_level[idx]
-        tile_spacing = wsi.spacings[tile_level]
-        tile_arr = wsi.get_patch(
-            self.x[idx],
-            self.y[idx],
-            self.tile_size_resized[idx],
-            self.tile_size_resized[idx],
-            spacing=tile_spacing,
-            center=False,
-        )
-        tile = Image.fromarray(tile_arr).convert("RGB")
-        if self.tile_size[idx] != self.tile_size_resized[idx]:
-            tile = tile.resize((self.tile_size[idx], self.tile_size[idx]))
-        if self.transforms:
-            if isinstance(self.transforms, BaseImageProcessor):  # Hugging Face (`transformer`) 
-                tile = self.transforms(tile, return_tensors="pt")["pixel_values"].squeeze(0)
-            else:  # general callable such as torchvision transforms
-                tile = self.transforms(tile)
-        return idx, tile