dataeval 1.1.3__tar.gz → 1.1.4__tar.gz

{dataeval-1.1.3 → dataeval-1.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: dataeval
-Version: 1.1.3
+Version: 1.1.4
 Summary: DataEval provides a simple interface to characterize image data and its impact on model performance across classification and object-detection tasks
 Project-URL: Homepage, https://dataeval.ai/
 Project-URL: Repository, https://github.com/aria-ml/dataeval/

{dataeval-1.1.3 → dataeval-1.1.4}/pyproject.toml

@@ -107,10 +107,6 @@ lint = [
 docsync = [
   "jupytext>=1.19.1",
 ]
-docker = [
-  "jinja2>=3.1",
-  "pyyaml>=6.0",
-]
 test = [
   "coverage[toml]>=7.6",
   "filelock>=3.20.3",

{dataeval-1.1.3 → dataeval-1.1.4}/src/dataeval/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '1.1.3'
-__version_tuple__ = version_tuple = (1, 1, 3)
+__version__ = version = '1.1.4'
+__version_tuple__ = version_tuple = (1, 1, 4)
 
 __commit_id__ = commit_id = None

{dataeval-1.1.3 → dataeval-1.1.4}/src/dataeval/extractors/_bovw.py

@@ -208,7 +208,9 @@ class BoVWExtractor:
             If no SIFT features are found in any image. This typically occurs
             when all images are uniform (e.g., solid color).
         """
-        with PoolWrapper(get_max_processes(), context="spawn") as pool:
+        # Threads, not processes: OpenCV and k-means release the GIL, so threads scale SIFT
+        # without paying process startup and model pickling on every Embeddings batch.
+        with PoolWrapper(get_max_processes(), context="threads") as pool:
             all_descriptors = [
                 des for _, des in sorted(pool.imap_unordered(_extract_single, enumerate(data))) if len(des) > 0
             ]
@@ -269,7 +271,7 @@ class BoVWExtractor:
 
         result = np.zeros((len(images), n_clusters), dtype=np.float32)
 
-        with PoolWrapper(get_max_processes(), context="spawn") as pool:
+        with PoolWrapper(get_max_processes(), context="threads") as pool:
             for idx, histogram in pool.imap_unordered(
                 partial(_transform_single, kmeans=self._kmeans, n_clusters=n_clusters),
                 enumerate(images),

{dataeval-1.1.3 → dataeval-1.1.4}/src/dataeval/utils/_internal.py

@@ -8,6 +8,7 @@ import sys
 from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence
 from enum import Enum
 from importlib import import_module
+from multiprocessing.pool import ThreadPool
 from os import cpu_count
 from types import ModuleType
 from typing import Any, Literal, TypeVar, cast, overload
@@ -948,9 +949,14 @@ class PoolWrapper:
     Also supports 'threads' for workloads where the GIL is released during computation.
     """
 
-    def __init__(self, processes: int | None, context: Literal["fork", "spawn"] = DEFAULT_CONTEXT) -> None:
+    def __init__(self, processes: int | None, context: Literal["fork", "spawn", "threads"] = DEFAULT_CONTEXT) -> None:
         procs = 1 if processes is None else max(1, (cpu_count() or 1) + processes + 1) if processes < 0 else processes
-        self._pool = multiprocessing.get_context(context).Pool(procs) if procs > 1 else None
+        if procs <= 1:
+            self._pool = None
+        elif context == "threads":
+            self._pool = ThreadPool(procs)
+        else:
+            self._pool = multiprocessing.get_context(context).Pool(procs)
 
     def imap_unordered(self, func: Callable[[T], R], iterable: Iterable[T]) -> Iterator[R]:
         """Apply `func` to each item in `iterable`, optionally using a pool."""