dataeval 0.86.9__tar.gz → 0.87.0__tar.gz

{dataeval-0.86.9 → dataeval-0.87.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dataeval
-Version: 0.86.9
+Version: 0.87.0
 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/
@@ -8,7 +8,7 @@ Project-URL: Documentation, https://dataeval.readthedocs.io/
 Author-email: Andrew Weng <andrew.weng@ariacoustics.com>, Bill Peria <bill.peria@ariacoustics.com>, Jon Botts <jonathan.botts@ariacoustics.com>, Jonathan Christian <jonathan.christian@ariacoustics.com>, Justin McMillan <justin.mcmillan@ariacoustics.com>, Ryan Wood <ryan.wood@ariacoustics.com>, Scott Swan <scott.swan@ariacoustics.com>, Shaun Jullens <shaun.jullens@ariacoustics.com>
 Maintainer-email: ARiA <dataeval@ariacoustics.com>
 License-Expression: MIT
-License-File: LICENSE.txt
+License-File: LICENSE
 Classifier: Development Status :: 4 - Beta
 Classifier: Intended Audience :: Science/Research
 Classifier: License :: OSI Approved :: MIT License
@@ -20,15 +20,12 @@ Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Topic :: Scientific/Engineering
 Requires-Python: <3.13,>=3.9
-Requires-Dist: defusedxml>=0.7.1
 Requires-Dist: fast-hdbscan==0.2.0
 Requires-Dist: lightgbm>=4
 Requires-Dist: numba>=0.59.1
 Requires-Dist: numpy>=1.24.2
 Requires-Dist: pandas>=2.0
-Requires-Dist: pillow>=10.3.0
 Requires-Dist: polars>=1.0.0
-Requires-Dist: requests>=2.32.3
 Requires-Dist: scikit-learn>=1.5.0
 Requires-Dist: scipy>=1.10
 Requires-Dist: torch>=2.2.0
@@ -123,14 +120,8 @@ micromamba create -f environment\environment.yaml -c pytorch
 
 ### **Installing from GitHub**
 
-To install DataEval from source locally on Ubuntu, you will need `git-lfs` to
-download larger, binary source files.
-
-```bash
-sudo apt-get install git-lfs
-```
-
-Pull the source down and change to the DataEval project directory.
+To install DataEval from source locally on Ubuntu, pull the source down and
+change to the DataEval project directory.
 
 ```bash
 git clone https://github.com/aria-ml/dataeval.git
@@ -167,10 +158,7 @@ source .venv/bin/activate
 
 ## Contact Us
 
-If you have any questions, feel free to reach out to the people below:
-
-- **POC**: Scott Swan @scott.swan
-- **DPOC**: Andrew Weng @aweng
+If you have any questions, feel free to reach out to [us](mailto:dataeval@ariacoustics.com)!
 
 ## Acknowledgement
 

{dataeval-0.86.9 → dataeval-0.87.0}/README.md

@@ -72,14 +72,8 @@ micromamba create -f environment\environment.yaml -c pytorch
 
 ### **Installing from GitHub**
 
-To install DataEval from source locally on Ubuntu, you will need `git-lfs` to
-download larger, binary source files.
-
-```bash
-sudo apt-get install git-lfs
-```
-
-Pull the source down and change to the DataEval project directory.
+To install DataEval from source locally on Ubuntu, pull the source down and
+change to the DataEval project directory.
 
 ```bash
 git clone https://github.com/aria-ml/dataeval.git
@@ -116,10 +110,7 @@ source .venv/bin/activate
 
 ## Contact Us
 
-If you have any questions, feel free to reach out to the people below:
-
-- **POC**: Scott Swan @scott.swan
-- **DPOC**: Andrew Weng @aweng
+If you have any questions, feel free to reach out to [us](mailto:dataeval@ariacoustics.com)!
 
 ## Acknowledgement
 

{dataeval-0.86.9 → dataeval-0.87.0}/pyproject.toml

@@ -31,15 +31,12 @@ classifiers = [
   "Topic :: Scientific/Engineering",
 ]
 dependencies = [
-  "defusedxml>=0.7.1",
   "fast_hdbscan==0.2.0",
   "lightgbm>=4",
   "numba>=0.59.1",
   "numpy>=1.24.2",
   "pandas>=2.0",
-  "pillow>=10.3.0",
   "polars>=1.0.0",
-  "requests>=2.32.3",
   "scipy>=1.10",
   "scikit-learn>=1.5.0",
   "torch>=2.2.0",
@@ -96,6 +93,7 @@ docs = [
   "jinja2>=3.1.6",
   "jupyter-client>=8.6.0",
   "jupyter-cache>=1.0",
+  "maite-datasets>=0.0.1",
   "myst-nb>=1.0",
   "sphinx-autoapi>=3.6.0",
   "sphinx-design>=0.6.1",

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/__init__.py

@@ -9,7 +9,7 @@ from __future__ import annotations
 
 try:
     from ._version import __version__
-except ImportError:
+except ImportError:  # pragma: no cover
     __version__ = "unknown"
 
 # Strongly type for pyright

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/_version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.86.9'
-__version_tuple__ = version_tuple = (0, 86, 9)
+__version__ = version = '0.87.0'
+__version_tuple__ = version_tuple = (0, 87, 0)

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/config.py

@@ -4,19 +4,15 @@ Global configuration settings for DataEval.
 
 from __future__ import annotations
 
-__all__ = ["get_device", "set_device", "get_max_processes", "set_max_processes", "use_max_processes", "DeviceLike"]
+__all__ = ["get_device", "set_device", "get_max_processes", "set_max_processes", "use_max_processes"]
 
-import sys
-from typing import Any, Union
-
-if sys.version_info >= (3, 10):
-    from typing import TypeAlias
-else:
-    from typing_extensions import TypeAlias
+from typing import Any
 
 import numpy as np
 import torch
 
+from dataeval.typing import DeviceLike
+
 ### GLOBALS ###
 
 _device: torch.device | None = None
@@ -27,17 +23,6 @@ _seed: int | None = None
 
 EPSILON = 1e-12
 
-### TYPES ###
-
-DeviceLike: TypeAlias = Union[int, str, tuple[str, int], torch.device]
-"""
-Type alias for types that are acceptable for specifying a torch.device.
-
-See Also
---------
-`torch.device <https://pytorch.org/docs/stable/tensor_attributes.html#torch.device>`_
-"""
-
 ### FUNCS ###
 
 

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/data/_metadata.py

@@ -16,18 +16,31 @@ from dataeval.typing import (
     ObjectDetectionTarget,
 )
 from dataeval.utils._array import as_numpy
-from dataeval.utils._bin import bin_data, digitize_data
+from dataeval.utils._bin import bin_data, digitize_data, is_continuous
 from dataeval.utils.data.metadata import merge
 
 
 def _binned(name: str) -> str:
-    return f"{name}[]"
+    return f"{name}↕"
+
+
+def _digitized(name: str) -> str:
+    return f"{name}#"
 
 
 @dataclass
 class FactorInfo:
-    factor_type: Literal["categorical", "continuous", "discrete"] | None = None
-    discretized_col: str | None = None
+    factor_type: Literal["categorical", "continuous", "discrete"]
+    is_binned: bool = False
+    is_digitized: bool = False
+
+
+def _to_col(name: str, info: FactorInfo, binned: bool = True) -> str:
+    if binned and info.is_binned:
+        return _binned(name)
+    if info.is_digitized:
+        return _digitized(name)
+    return name
 
 
 class Metadata:
@@ -60,7 +73,7 @@ class Metadata:
         self._class_labels: NDArray[np.intp]
         self._class_names: list[str]
         self._image_indices: NDArray[np.intp]
-        self._factors: dict[str, FactorInfo]
+        self._factors: dict[str, FactorInfo | None]
         self._dropped_factors: dict[str, list[str]]
         self._dataframe: pl.DataFrame
         self._raw: Sequence[Mapping[str, Any]]
@@ -146,14 +159,27 @@ class Metadata:
         return self._dropped_factors
 
     @property
-    def discretized_data(self) -> NDArray[np.int64]:
-        """Factor data with continuous data discretized."""
+    def digitized_data(self) -> NDArray[np.int64]:
+        """Factor data with digitized categorical data."""
+        if not self.factor_names:
+            return np.array([], dtype=np.int64)
+
+        self._bin()
+        return (
+            self.dataframe.select([_to_col(k, v, False) for k, v in self.factor_info.items()])
+            .to_numpy()
+            .astype(np.int64)
+        )
+
+    @property
+    def binned_data(self) -> NDArray[np.int64]:
+        """Factor data with binned continuous data."""
         if not self.factor_names:
             return np.array([], dtype=np.int64)
 
         self._bin()
         return (
-            self.dataframe.select([info.discretized_col or name for name, info in self.factor_info.items()])
+            self.dataframe.select([_to_col(k, v, True) for k, v in self.factor_info.items()])
             .to_numpy()
             .astype(np.int64)
         )
@@ -168,7 +194,7 @@ class Metadata:
     def factor_info(self) -> Mapping[str, FactorInfo]:
         """Factor types of the metadata."""
         self._bin()
-        return dict(filter(self._filter, self._factors.items()))
+        return dict(filter(self._filter, ((k, v) for k, v in self._factors.items() if v is not None)))
 
     @property
     def factor_data(self) -> NDArray[Any]:
@@ -194,7 +220,7 @@ class Metadata:
     @property
     def image_indices(self) -> NDArray[np.intp]:
         """Indices of images as a NumPy array."""
-        self._bin()
+        self._structure()
         return self._image_indices
 
     @property
@@ -212,7 +238,7 @@ class Metadata:
             columns = self._dataframe.columns
             for col in (col for col in cols or columns if _binned(col) in columns):
                 self._dataframe.drop_in_place(_binned(col))
-                self._factors[col] = FactorInfo()
+                self._factors[col] = None
             self._is_binned = False
 
     def _structure(self) -> None:
@@ -277,7 +303,7 @@ class Metadata:
         self._class_labels = labels
         self._class_names = list(index2label.values())
         self._image_indices = target_dict["image_index"]
-        self._factors = dict.fromkeys(factor_dict, FactorInfo())
+        self._factors = dict.fromkeys(factor_dict, None)
         self._dataframe = pl.DataFrame({**target_dict, **factor_dict})
         self._dropped_factors = merged[1]
         self._is_structured = True
@@ -303,24 +329,25 @@ class Metadata:
             )
 
         column_set = set(df.columns)
-        for col in (col for col in self.factor_names if _binned(col) not in column_set):
+        for col in (col for col in self.factor_names if not {_binned(col), _digitized(col)} & column_set):
             # Get data as numpy array for processing
             data = df[col].to_numpy()
-            col_dz = _binned(col)
             if col in factor_bins:
                 # User provided binning
                 bins = factor_bins[col]
-                df = df.with_columns(pl.Series(name=col_dz, values=digitize_data(data, bins).astype(np.int64)))
-                factor_info[col] = FactorInfo("continuous", col_dz)
+                col_bn = _binned(col)
+                df = df.with_columns(pl.Series(name=col_bn, values=digitize_data(data, bins).astype(np.int64)))
+                factor_info[col] = FactorInfo("continuous", is_binned=True)
             else:
                 # Check if data is numeric
-                unique, ordinal = np.unique(data, return_inverse=True)
-                if not np.issubdtype(data.dtype, np.number) or unique.size <= max(20, data.size * 0.01):
-                    # Non-numeric data or small number of unique values - convert to categorical
-                    df = df.with_columns(pl.Series(name=col_dz, values=ordinal.astype(np.int64)))
-                    factor_info[col] = FactorInfo("categorical", col_dz)
-                elif data.dtype == float:
-                    # Many unique values - discretize by binning
+                _, ordinal = np.unique(data, return_inverse=True)
+                if not np.issubdtype(data.dtype, np.number):
+                    # Non-numeric data - convert to categorical
+                    col_dg = _digitized(col)
+                    df = df.with_columns(pl.Series(name=col_dg, values=ordinal.astype(np.int64)))
+                    factor_info[col] = FactorInfo("categorical", is_digitized=True)
+                elif is_continuous(data, self.image_indices):
+                    # Continuous values - discretize by binning
                     warnings.warn(
                         f"A user defined binning was not provided for {col}. "
                         f"Using the {self.auto_bin_method} method to discretize the data. "
@@ -330,10 +357,12 @@ class Metadata:
                     )
                     # Create binned version
                     binned_data = bin_data(data, self.auto_bin_method)
-                    df = df.with_columns(pl.Series(name=col_dz, values=binned_data.astype(np.int64)))
-                    factor_info[col] = FactorInfo("continuous", col_dz)
+                    col_bn = _binned(col)
+                    df = df.with_columns(pl.Series(name=col_bn, values=binned_data.astype(np.int64)))
+                    factor_info[col] = FactorInfo("continuous", is_binned=True)
                 else:
-                    factor_info[col] = FactorInfo("discrete", col)
+                    # Non-continuous values - treat as discrete
+                    factor_info[col] = FactorInfo("discrete")
 
         # Store the results
         self._dataframe = df
@@ -367,7 +396,7 @@ class Metadata:
         for k, v in factors.items():
             data = as_numpy(v)[self.image_indices]
             new_columns.append(pl.Series(name=k, values=data))
-            self._factors[k] = FactorInfo()
+            self._factors[k] = None
 
         if new_columns:
             self._dataframe = self.dataframe.with_columns(new_columns)

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/data/_split.py

@@ -208,7 +208,7 @@ def get_groups(metadata: Metadata, split_on: Sequence[str] | None) -> NDArray[np
 
     split_set = set(split_on)
     indices = [i for i, name in enumerate(metadata.factor_names) if name in split_set]
-    binned_features = metadata.discretized_data[:, indices]
+    binned_features = metadata.binned_data[:, indices]
     return np.unique(binned_features, axis=0, return_inverse=True)[1]
 
 

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/data/selections/_classbalance.py

@@ -11,12 +11,13 @@ from dataeval.utils._array import as_numpy
 
 class ClassBalance(Selection[ImageClassificationDatum]):
     """
-    Balance the dataset by class.
+    Select indices of a dataset that will equalize the occurrences of all classes.
 
     Note
     ----
-    The total number of instances of each class will be equalized which may result
+    1. The total number of instances of each class will be equalized which may result
     in a lower total number of instances than specified by the selection limit.
+    2. This selection currently only supports classification tasks
     """
 
     stage = SelectionStage.FILTER
@@ -29,7 +30,7 @@ class ClassBalance(Selection[ImageClassificationDatum]):
                 label = int(np.argmax(as_numpy(target)))
             else:
                 # ObjectDetectionTarget and SegmentationTarget not supported yet
-                raise TypeError("ClassFilter only supports classification targets as an array of confidence scores.")
+                raise TypeError("ClassBalance only supports classification targets as an array of class probabilities.")
             class_indices.setdefault(label, []).append(i)
 
         per_class_limit = min(min(len(c) for c in class_indices.values()), dataset._size_limit // len(class_indices))

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/data/selections/_classfilter.py

@@ -14,12 +14,12 @@ from dataeval.utils._array import as_numpy
 
 class ClassFilter(Selection[Any]):
     """
-    Filter the dataset by class.
+    Select dataset indices based on class labels, keeping only those present in `classes`.
 
     Parameters
     ----------
     classes : Sequence[int]
-        The classes to filter by.
+        The sequence of classes to keep.
     filter_detections : bool, default True
         Whether to filter detections from targets for object detection and segmentation datasets.
     """
@@ -41,16 +41,16 @@ class ClassFilter(Selection[Any]):
             if isinstance(target, Array):
                 # Get the label for the image
                 label = int(np.argmax(as_numpy(target)))
-                # Check to see if the label is in the classes to filter for
+                # Check to see if the label is in the classes to keep
                 if label in self.classes:
-                    # Include the image
+                    # Include the image index
                     selection.append(idx)
             elif isinstance(target, (ObjectDetectionTarget, SegmentationTarget)):
                 # Get the set of labels from the target
                 labels = set(target.labels if isinstance(target.labels, Iterable) else [target.labels])
                 # Check to see if any labels are in the classes to filter for
                 if labels.intersection(self.classes):
-                    # Include the image
+                    # Include the image index
                     selection.append(idx)
                     # If we are filtering out other labels and there are other labels, add a subselection filter
                     if self.filter_detections and labels.difference(self.classes):

{dataeval-0.86.9 → dataeval-0.87.0}/src/dataeval/data/selections/_indices.py

@@ -9,12 +9,12 @@ from dataeval.data._selection import Select, Selection, SelectionStage
 
 class Indices(Selection[Any]):
     """
-    Selects specific indices from the dataset.
+    Selects only the given indices from the dataset.
 
     Parameters
     ----------
     indices : Sequence[int]
-        The indices to select from the dataset.
+        The specific indices to select.
     """
 
     stage = SelectionStage.FILTER