datatunner 3.2.1__tar.gz → 3.2.2__tar.gz

{datatunner-3.2.1/datatunner.egg-info → datatunner-3.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: datatunner
-Version: 3.2.1
+Version: 3.2.2
 Summary: DataTunner: Scientific Platform for Optimal Artificial Data Proportion in Deep Learning
 Author-email: Leandro Costa Rocha <leandrocrx@gmail.com>
 License-Expression: MIT

{datatunner-3.2.1 → datatunner-3.2.2}/datatunner/orchestrator.py

@@ -236,6 +236,18 @@ class DataTunner:
 
         return report
 
+    @staticmethod
+    def _count_samples(data: Any) -> int:
+        """Number of samples in a dataset, handling (images, labels) tuples.
+
+        Image pipelines pass real_data as a (X, y) tuple; ``len()`` on the
+        tuple returns the number of elements (2), not the number of samples.
+        Without this, n_synthetic = int(alpha * 2) ≈ 0 for every alpha.
+        """
+        if isinstance(data, tuple) and len(data) == 2:
+            return len(data[0])
+        return len(data)
+
     def _run_single_experiment(self, alpha: float, seed: int) -> ExperimentResult:
         """Execute ONE complete experiment: generate -> mix -> train -> evaluate.
 
@@ -245,7 +257,7 @@ class DataTunner:
 
         with self.environment.isolate(seed, experiment_id):
             # 1. Generate synthetic data
-            n_real = len(self._real_data)
+            n_real = self._count_samples(self._real_data)
             n_synthetic = int(alpha * n_real)
 
             # Bind the generator RNG to this trial's seed so repetitions
@@ -346,7 +358,7 @@ class DataTunner:
     ) -> List[Dict]:
         """Pre-evaluate generator by generating a sample and measuring fidelity."""
         # Generate a fixed-size sample for quality assessment
-        sample_size = min(1000, len(real_data))
+        sample_size = min(1000, self._count_samples(real_data))
         generator.set_seed(generator.spec.random_state)
         synthetic_sample = generator.generate(sample_size)
 

{datatunner-3.2.1 → datatunner-3.2.2/datatunner.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: datatunner
-Version: 3.2.1
+Version: 3.2.2
 Summary: DataTunner: Scientific Platform for Optimal Artificial Data Proportion in Deep Learning
 Author-email: Leandro Costa Rocha <leandrocrx@gmail.com>
 License-Expression: MIT

{datatunner-3.2.1 → datatunner-3.2.2}/datatunner.egg-info/SOURCES.txt

@@ -64,6 +64,7 @@ tests/__init__.py
 tests/test_aggregation.py
 tests/test_domain.py
 tests/test_generators.py
+tests/test_image_support.py
 tests/test_infrastructure.py
 tests/test_metrics_formatting.py
 tests/test_mixing.py

{datatunner-3.2.1 → datatunner-3.2.2}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "datatunner"
-# Bump on release: the current published version on PyPI is 3.1.1.
-version = "3.2.1"
+# Bump on release: the current published version on PyPI is 3.2.1.
+version = "3.2.2"
 description = "DataTunner: Scientific Platform for Optimal Artificial Data Proportion in Deep Learning"
 readme = "README.md"
 license = "MIT"

datatunner-3.2.2/tests/test_image_support.py

@@ -0,0 +1,39 @@
+"""Regression tests for image pipeline support (X, y) tuples.
+
+Bug fix (2026-08): `_run_single_experiment` computed n_synthetic with
+``len(real_data)``. For image pipelines real_data is a (X, y) tuple, so
+len() returned 2 → n_synthetic = int(alpha * 2) ≈ 0 for every alpha:
+the ImageAugmentation pipeline generated (almost) no synthetic data.
+
+Fix: DataTunner._count_samples() counts rows of data[0] for 2-tuples.
+Used in _run_single_experiment and _evaluate_generator_quality.
+
+Run where deps are installed (CI / Colab):  pytest tests/test_image_support.py
+"""
+import pytest
+
+from datatunner.orchestrator import DataTunner
+
+
+class TestCountSamples:
+    @pytest.mark.parametrize("n", [0, 1, 2, 10, 1000])
+    def test_tuple_counts_samples_not_elements(self, n):
+        X = [0] * n
+        y = [0] * n
+        assert DataTunner._count_samples((X, y)) == n
+
+    def test_nested_2d_like_image_array(self):
+        # X shaped (N, H, W): must count N, not dimensions
+        X = [[[0] * 4 for _ in range(4)] for _ in range(250)]
+        y = [0] * 250
+        assert DataTunner._count_samples((X, y)) == 250
+
+    def test_plain_list_counts_len(self):
+        assert DataTunner._count_samples([1, 2, 3, 4]) == 4
+
+    def test_single_element_list(self):
+        assert DataTunner._count_samples(["only"]) == 1
+
+    def test_non_tuple_pair_not_misclassified(self):
+        # A 2-element tuple is ambiguous; contract: (X, y) tuples mean images
+        assert DataTunner._count_samples((1, 2)) == 1  # len(data[0]) = 1