Merge branch 'ls-release/1.17.0' of gitcode.com:ModelEngine/Label-Studio into ls-release/1.17.0

Jason Wang · Jason Wang · commit 3ee6cfcc6a8b · 2025-06-03T19:07:00.000+08:00
diff --git a/.dockerignore b/.dockerignore
@@ -8,6 +8,7 @@
 !deploy/uwsgi.ini
 !deploy/default.conf
 !label_studio/**
+!wheels/**
 # Re-ignore pyc files and frontend
 label_studio/frontend
 label_studio/**/*.pyc
diff --git a/Dockerfile b/Dockerfile
@@ -74,11 +74,13 @@ RUN --mount=type=cache,target="/var/cache/apt",sharing=locked \
     set -eux; \
     apt-get update; \
     apt-get install --no-install-recommends -y \
-            build-essential git; \
+    build-essential git; \
     apt-get autoremove -y
 
 WORKDIR /label-studio
 
+COPY wheels/ wheels/
+
 ENV VENV_PATH="/label-studio/.venv"
 ENV PATH="$VENV_PATH/bin:$PATH"
 
@@ -94,11 +96,14 @@ ARG INCLUDE_DEV=false
 RUN --mount=type=cache,target=$POETRY_CACHE_DIR,sharing=locked \
     poetry check --lock && \
     if [ "$INCLUDE_DEV" = "true" ]; then \
-        poetry install --no-root --extras uwsgi --with test; \
+    poetry install --no-root --extras uwsgi --with test; \
     else \
-        poetry install --no-root --without test --extras uwsgi; \
+    poetry install --no-root --without test --extras uwsgi; \
     fi
 
+# Install extra Python packages (from wheels and PyPI)
+RUN pip install wheels/*.whl openslide-bin openslide-python loguru
+
 # Install LS
 COPY label_studio label_studio
 RUN --mount=type=cache,target=$POETRY_CACHE_DIR,sharing=locked \
@@ -135,8 +140,9 @@ RUN --mount=type=cache,target="/var/cache/apt",sharing=locked \
     set -eux; \
     apt-get update; \
     apt-get upgrade -y; \
-    apt-get install --no-install-recommends -y libexpat1 \
-        gnupg2 curl; \
+    apt-get install --no-install-recommends -y \
+        libexpat1 gnupg2 curl \
+        libgl1 libopenjp2-7 libnuma1 libglib2.0-0 libspeex1 libvdpau1 libgomp1; \
     apt-get autoremove -y
 
 # install nginx
diff --git a/label_studio/core/annotated_deepzoom_generator.py b/label_studio/core/annotated_deepzoom_generator.py
@@ -0,0 +1,155 @@
+from __future__ import annotations
+from collections.abc import Callable
+import os
+from pathlib import Path
+
+import cv2
+import openslide
+from openslide.deepzoom import DeepZoomGenerator
+from typing import TYPE_CHECKING
+from PIL import Image
+import numpy as np
+import math
+
+if TYPE_CHECKING:
+    from typing import TypeAlias
+
+    Transform: TypeAlias = Callable[[Image.Image], None]
+
+ENABLE_DEBUG = False
+
+
+class AnnotatedDeepZoomGenerator(DeepZoomGenerator):
+    filename: str
+    full_path: Path
+    mpp: float
+    transform: Transform
+
+    def __init__(
+        self,
+        osr,
+        full_path: Path,
+        tile_size: int = 254,
+        overlap: int = 1,
+        limit_bounds: bool = False,
+    ):
+        super().__init__(osr, tile_size, overlap, limit_bounds)
+        # heatmap = full_path.with_name(f"{full_path.stem}.heatmap.npy")
+        self.is_sdpc = full_path.suffix == ".sdpc"
+        # print(f"[HEATMAP] (exist: {os.path.isfile(heatmap)}) fetching heatmap: ", heatmap)
+        self.visited = set()
+        # if os.path.isfile(heatmap):
+        #     self.color_mask = np.load(heatmap)
+
+        #     self.actual_size = self.color_mask.shape
+        #     gap = [abs(dim[0] - self.actual_size[1]) for dim in self._osr.level_dimensions]
+        #     self.ratio = gap.index(min(gap))
+        #     self.region_size = self._osr.level_dimensions[self.ratio]
+        #     largest = self._osr.level_dimensions[0]
+        #     # down_sample = self._osr.level_downsamples[self.ratio]
+        #     down_sample = 1 << round(math.log2(largest[0] / self.actual_size[1]))
+        #     offset = (
+        #         largest[0] / down_sample - self.color_mask.shape[1],
+        #         largest[1] / down_sample - self.color_mask.shape[0]
+        #     )
+        #     print(f"valid: {0 if any(offset) else 1} offset: ", offset)
+        # else:
+        #     self.color_mask = None
+        #     self.ratio = None
+        #     self.region_size = None
+
+    def get_tile(
+        self, level: int, address: tuple[int, int], heatmap=False
+    ) -> Image.Image:
+        """Return an RGB PIL.Image for a tile.
+
+        level:     the Deep Zoom level.
+        address:   the address of the tile within the level as a (col, row)
+                   tuple.
+        heatmap:   need heatmap overlay"""
+        args, z_size = self._get_tile_info(level, address)
+
+        if ENABLE_DEBUG and level not in self.visited:
+            print("read_regoin: ", args)
+
+        tile = self._osr.read_region(*args)
+        profile = tile.info.get("icc_profile")
+
+        # Apply on solid background
+        if isinstance(self._osr, openslide.OpenSlide):
+            bg = Image.new("RGB", tile.size, self._bg_color)
+            tile = Image.composite(tile, bg, tile)
+        if heatmap and self.color_mask is not None:
+            tile = Image.fromarray(self._mask_tile(tile, *args))
+
+        # Scale to the correct size
+        if tile.size != z_size:
+            # Image.Resampling added in Pillow 9.1.0
+            # Image.LANCZOS removed in Pillow 10
+            tile.thumbnail(z_size, getattr(Image, "Resampling", Image).LANCZOS)
+
+        # Reference ICC profile
+        if profile is not None:
+            tile.info["icc_profile"] = profile
+
+        return tile
+
+    def _mask_tile(
+        self, tile: Image, location: tuple[int, int], level: int, size: tuple[int, int]
+    ) -> np.ndarray:
+        region_size = (self.actual_size[1], self.actual_size[0])
+        if self.is_sdpc:
+            # down_sample = self._osr.level_downsamples[self.ratio]
+            down_sample = 1 << round(math.log2(self._osr.level_dimensions[0][0] / self.actual_size[1]))
+            lv_downsample = self._osr.level_downsamples[level]
+            scale = (lv_downsample / down_sample, lv_downsample / down_sample)
+            max_scale = (1 / down_sample, 1 / down_sample)
+        else:
+            lv_dim = self._osr.level_dimensions[level]
+            max_dim = self._osr.level_dimensions[0]
+            scale = (region_size[0] / lv_dim[0], region_size[1] / lv_dim[1])
+            max_scale = (region_size[0] / max_dim[0], region_size[1] / max_dim[1])
+        x_img = int(location[0] * max_scale[0])
+        y_img = int(location[1] * max_scale[1])
+        x_end = min(int(x_img + size[0] * scale[0]), region_size[0])
+        y_end = min(int(y_img + size[1] * scale[1]), region_size[1])
+
+        tile_arr = np.array(tile.convert("RGB"))
+
+        color_mask = self.color_mask[y_img:y_end, x_img:x_end]
+        
+        if ENABLE_DEBUG and level not in self.visited:
+            print("max_scale", max_scale)
+            print("lv: ", level)
+            print("region_size: ", region_size)
+            print("color_mask", color_mask.shape)
+            print("ori_color_mask", self.color_mask.shape)
+            print("size", size)
+            print("mask_size", color_mask.shape)
+            print("lv_downsample", self._osr.level_downsamples)
+            print("lv_dimension", self._osr.level_dimensions)
+            print("get_best_level_for_downsample ", self._osr.get_best_level_for_downsample(self.ratio))
+            print(f"taking mask ({x_img}, {y_img}) with len ({color_mask.shape[1]}, {color_mask.shape[0]})")
+            self.visited.add(level)
+        if color_mask.shape[0] == 0 or color_mask.shape[1] == 0:
+            return tile_arr
+        # color_mask = self._resize_image(color_mask, (size[1], size[0]))
+        color_mask = cv2.resize(color_mask, size)
+
+        weighted = color_mask * 0.4 + tile_arr * 0.6
+        ret = np.clip(weighted, 0, 255).astype(np.uint8)
+        return ret
+
+    def _resize_image(self, image, new_size):
+        resized_image = np.zeros(
+            (new_size[0], new_size[1], image.shape[2]), dtype=image.dtype
+        )
+        row_scale = image.shape[0] / new_size[0]
+        col_scale = image.shape[1] / new_size[1]
+
+        for j in range(new_size[1]):
+            for i in range(new_size[0]):
+                src_row = int(i * row_scale)
+                src_col = int(j * col_scale)
+                resized_image[i, j] = image[src_row, src_col]
+        return resized_image
diff --git a/label_studio/core/deepzoom_util.py b/label_studio/core/deepzoom_util.py
@@ -0,0 +1,39 @@
+import os
+from pathlib import Path
+
+import openslide
+import opensdpc
+
+from .annotated_deepzoom_generator import AnnotatedDeepZoomGenerator
+
+class DeepZoomWrapper:
+    def __init__(self, full_path, tile_size:int = 254, overlap:int = 1, limit_bounds:bool = False):
+        full_path = str(full_path)
+        _, ext = os.path.splitext(full_path)
+
+        if ext == ".sdpc":
+            self._osr = opensdpc.OpenSdpc(full_path)
+        else:
+            self._osr = openslide.OpenSlide(full_path)
+
+        self._dzg = AnnotatedDeepZoomGenerator(
+            self._osr,
+            full_path=Path(full_path),
+            tile_size=tile_size,
+            overlap=overlap,
+            limit_bounds=limit_bounds,
+        )
+
+    def get_tile(self, level: int, tile: tuple[int, int]):
+        return self._dzg.get_tile(level, tile)
+    
+    def get_dzi(self, format:str = "jpeg"):
+        return self._dzg.get_dzi(format)
+    
+    @property
+    def level_dimensions(self):
+        return self._dzg.level_dimensions
+    
+    @property
+    def level_count(self):
+        return self._dzg.level_count
diff --git a/label_studio/core/views.py b/label_studio/core/views.py
@@ -8,6 +8,7 @@
 import posixpath
 from pathlib import Path
 from wsgiref.util import FileWrapper
+from io import BytesIO
 
 import pandas as pd
 import requests
@@ -36,6 +37,8 @@
 from rest_framework.permissions import IsAuthenticated
 from rest_framework.views import APIView
 
+from .deepzoom_util import DeepZoomWrapper
+
 logger = logging.getLogger(__name__)
 
 
@@ -200,6 +203,11 @@ def localfiles_data(request):
     """Serving files for LocalFilesImportStorage"""
     user = request.user
     path = request.GET.get('d')
+
+    level = request.GET.get('level')
+    col = request.GET.get('col')
+    row = request.GET.get('row')
+
     if settings.LOCAL_FILES_SERVING_ENABLED is False:
         return HttpResponseForbidden(
             "Serving local files can be dangerous, so it's disabled by default. "
@@ -211,17 +219,55 @@ def localfiles_data(request):
     if path and request.user.is_authenticated:
         path = posixpath.normpath(path).lstrip('/')
         full_path = Path(safe_join(local_serving_document_root, path))
-        user_has_permissions = False
+
+        # Check if the file is a WSI file
+        ext = os.path.splitext(full_path)[1].lower()
+        is_wsi = ext in ['.svs', '.sdpc', '.tif', '.tiff', '.csp', '.kfb']
 
         # Try to find Local File Storage connection based prefix:
         # storage.path=/home/user, full_path=/home/user/a/b/c/1.jpg =>
         # full_path.startswith(path) => True
         localfiles_storage = LocalFilesImportStorage.objects.annotate(
             _full_path=Value(os.path.dirname(full_path), output_field=CharField())
         ).filter(_full_path__startswith=F('path'))
+
+        # Check if user has permissions to access this storage
+        user_has_permissions = False
         if localfiles_storage.exists():
             user_has_permissions = any(storage.project.has_permission(user) for storage in localfiles_storage)
 
+        if not user_has_permissions or not os.path.exists(full_path):
+            return HttpResponseNotFound()
+        
+        # Check if the file is a WSI file and has level, col, and row parameters
+        if is_wsi and level and col and row:
+            try:
+                level = int(level)
+                col = int(col)
+                row = int(row)
+            except ValueError:
+                return HttpResponseForbidden('Invalid level, col, or row parameter')
+
+            # Get the tile image
+            dz = DeepZoomWrapper(full_path)
+            tile_image = dz.get_tile(level, (col, row))
+
+            # Create a response with the tile image
+            buf = BytesIO()
+            tile_image.save(buf, 'jpeg')
+            buf.seek(0)
+            return HttpResponse(buf, content_type='image/jpeg')
+        
+        # Check if the file is a WSI file without level, col, and row parameters
+        if is_wsi:
+            # Create a DeepZoomWrapper instance
+            dz = DeepZoomWrapper(full_path)
+
+            # Get the DZI (Deep Zoom Image) for the WSI file
+            dzi_xml = dz.get_dzi(format='jpeg')
+            return HttpResponse(dzi_xml, content_type='application/xml')
+        
+        # If the file is not a WSI file, serve it as a regular file
         if user_has_permissions and os.path.exists(full_path):
             content_type, encoding = mimetypes.guess_type(str(full_path))
             content_type = content_type or 'application/octet-stream'
