sampledisco 0.4.0__tar.gz → 0.4.2__tar.gz

{sampledisco-0.4.0/src/sampledisco.egg-info → sampledisco-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sampledisco
-Version: 0.4.0
+Version: 0.4.2
 Summary: Cross-omics, cross-condition sample embedding for single-cell data (RNA + ATAC)
 Author: Hongkai Ji
 Author-email: Harry Jiang <hjiang55@jh.edu>
@@ -21,6 +21,7 @@ Requires-Dist: numpy>=1.23
 Requires-Dist: pandas>=1.5
 Requires-Dist: scipy>=1.9
 Requires-Dist: scikit-learn>=1.2
+Requires-Dist: threadpoolctl>=3.1
 Requires-Dist: scikit-image>=0.19
 Requires-Dist: scikit-misc>=0.3
 Requires-Dist: statsmodels>=0.13

{sampledisco-0.4.0 → sampledisco-0.4.2}/pyproject.toml

@@ -9,7 +9,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sampledisco"
-version = "0.4.0"
+version = "0.4.2"
 description = "Cross-omics, cross-condition sample embedding for single-cell data (RNA + ATAC)"
 readme = "README.md"
 requires-python = ">=3.10"

{sampledisco-0.4.0 → sampledisco-0.4.2}/requirements.txt

@@ -7,6 +7,7 @@ numpy>=1.23
 pandas>=1.5
 scipy>=1.9
 scikit-learn>=1.2
+threadpoolctl>=3.1   # imported directly (ThreadpoolController)
 scikit-image>=0.19   # scrublet-based ATAC doublet detection (doublet_detection=True)
 scikit-misc>=0.3
 statsmodels>=0.13

{sampledisco-0.4.0 → sampledisco-0.4.2}/src/sampledisco/parameter_selection/autotune.py

@@ -10,6 +10,9 @@ the data can't support:
     (CCA, SPS, CV-kNN, pseudotime-Spearman).
   - If neither → emit a warning and short-circuit with fixed defaults.
 
+``scoring="ilisi_label"`` is the two-term variant: one biology term (grouping
+tracking: CCA or categorical PC-R²) and one batch term (iLISI), equally weighted.
+
 Generalized version of `wire_autotune_dualembed_v2.py`. No dataset-specific
 paths; all data flows through the same `compute_sample_embedding` primitives
 in `sample_embedding/blocks.py`.
@@ -427,17 +430,20 @@ def make_scorer(name: str, meta: Dict, lam: float = 0.5) -> Callable[[np.ndarray
             - lam * (_asw_safe(emb, batch) if has_batch else 0.0)
         )
 
-    if name in ("multi_metric_proxy", "auto"):
+    if name in ("multi_metric_proxy", "auto", "ilisi_label"):
+        two_term = name == "ilisi_label"
         components: List[Callable[[np.ndarray], float]] = []
         if has_grouping:
             tracking_key = "pc_r2_categorical" if grouping_is_categorical else "cca"
             components.append(
                 lambda emb: _minmax(_grouping_tracking_score(emb),
                                     *SCORING_BOUNDS[tracking_key]))
-            components.append(lambda emb: _minmax(_sps_continuous(emb, grouping), *SCORING_BOUNDS["sps"]))
+            if not two_term:
+                components.append(lambda emb: _minmax(_sps_continuous(emb, grouping), *SCORING_BOUNDS["sps"]))
         if has_batch:
             components.append(lambda emb: _minmax(_ilisi_norm(emb, batch), *SCORING_BOUNDS["ilisi_norm"]))
-            components.append(lambda emb: _minmax(-_asw_safe(emb, batch), *SCORING_BOUNDS["neg_asw_batch"]))
+            if not two_term:
+                components.append(lambda emb: _minmax(-_asw_safe(emb, batch), *SCORING_BOUNDS["neg_asw_batch"]))
         if not components:
             return lambda emb: 0.0
 
@@ -760,12 +766,13 @@ _PROXY_DESCRIPTIONS = {
 
 def _active_proxies(scoring: str, has_batch: bool, has_grouping: bool):
     """Return the list of proxy names the scorer actually evaluates."""
-    if scoring in ("auto", "multi_metric_proxy"):
+    if scoring in ("auto", "multi_metric_proxy", "ilisi_label"):
+        two_term = scoring == "ilisi_label"
         names = []
         if has_grouping:
-            names += ["cca", "sps"]
+            names += ["cca"] if two_term else ["cca", "sps"]
         if has_batch:
-            names += ["ilisi_batch", "neg_asw_batch"]
+            names += ["ilisi_batch"] if two_term else ["ilisi_batch", "neg_asw_batch"]
         return names
     return [scoring] if scoring in _PROXY_DESCRIPTIONS else [scoring]
 
@@ -818,8 +825,8 @@ def _format_autotune_report(*, best_params, best_score, trace, weights,
         for name in proxies:
             kind, desc = _PROXY_DESCRIPTIONS.get(name, ("?", name))
             lines.append(f"  - [{kind:<12s}] {name:<20s}  {desc}")
-        if scoring in ("auto", "multi_metric_proxy"):
-            lines.append("  ensemble: multi_metric_proxy = mean of the proxies above (each min-max scaled).")
+        if scoring in ("auto", "multi_metric_proxy", "ilisi_label"):
+            lines.append(f"  ensemble: {scoring} = mean of the proxies above (each min-max scaled).")
     lines.append("")
     lines.append("Result")
     lines.append("-" * 68)
@@ -928,7 +935,7 @@ def _finalize(adata, blocks, final_emb, weights, *,
             tune_on_modality=tune_on_modality, score_n_units=score_n_units,
         )
         report_path = os.path.join(out_dir, "autotune_record.txt")
-        with open(report_path, "w") as f:
+        with open(report_path, "w", encoding="utf-8") as f:
             f.write(report)
         if verbose:
             print(f"[autotune] wrote {emb_csv}")

{sampledisco-0.4.0 → sampledisco-0.4.2}/src/sampledisco/sample_embedding/blocks.py

@@ -152,16 +152,22 @@ def kmeans_centers_pair(Z: np.ndarray, K_med: int, K_fine: int, seed: int,
     """Centres at K_med (seed) and K_fine (seed + 1). With n_threads > 1 the two independent
     fits run concurrently (each keeps its own RandomState). BLAS (process-wide) is held at one
     thread because k-means++ issues thousands of tiny GEMMs; the OpenMP thread count is a
-    per-thread setting, so each worker sets its own share."""
+    per-thread setting, so each worker sets its own share.
+
+    sklearn.cluster is imported here, in the calling thread, so no library is loaded inside the
+    worker threads: a library scan (threadpoolctl) in one thread racing a first import of a
+    compiled extension in another can deadlock."""
     if n_threads < 2:
         return kmeans_centers(Z, K_med, seed), kmeans_centers(Z, K_fine, seed + 1)
-    from threadpoolctl import threadpool_limits
+    from sklearn.cluster import MiniBatchKMeans  # noqa: F401  (load sklearn's extensions before the threads start)
+    from threadpoolctl import ThreadpoolController
+    ctl = ThreadpoolController()
 
     def fit(K, sd):
-        with threadpool_limits(limits=max(1, n_threads // 2), user_api="openmp"):
+        with ctl.limit(limits=max(1, n_threads // 2), user_api="openmp"):
             return kmeans_centers(Z, K, sd)
 
-    with threadpool_limits(limits=1, user_api="blas"), ThreadPoolExecutor(2) as ex:
+    with ctl.limit(limits=1, user_api="blas"), ThreadPoolExecutor(2) as ex:
         f_med = ex.submit(fit, K_med, seed)
         f_fine = ex.submit(fit, K_fine, seed + 1)
         return f_med.result(), f_fine.result()

{sampledisco-0.4.0 → sampledisco-0.4.2/src/sampledisco.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sampledisco
-Version: 0.4.0
+Version: 0.4.2
 Summary: Cross-omics, cross-condition sample embedding for single-cell data (RNA + ATAC)
 Author: Hongkai Ji
 Author-email: Harry Jiang <hjiang55@jh.edu>
@@ -21,6 +21,7 @@ Requires-Dist: numpy>=1.23
 Requires-Dist: pandas>=1.5
 Requires-Dist: scipy>=1.9
 Requires-Dist: scikit-learn>=1.2
+Requires-Dist: threadpoolctl>=3.1
 Requires-Dist: scikit-image>=0.19
 Requires-Dist: scikit-misc>=0.3
 Requires-Dist: statsmodels>=0.13

{sampledisco-0.4.0 → sampledisco-0.4.2}/src/sampledisco.egg-info/requires.txt

@@ -2,6 +2,7 @@ numpy>=1.23
 pandas>=1.5
 scipy>=1.9
 scikit-learn>=1.2
+threadpoolctl>=3.1
 scikit-image>=0.19
 scikit-misc>=0.3
 statsmodels>=0.13

{sampledisco-0.4.0 → sampledisco-0.4.2}/tests/test_sample_embedding_equivalence.py

@@ -2,6 +2,8 @@
 save_cell_adata switch leaves the cell-level h5ad untouched. Synthetic data, CPU only.
 Run with `pytest tests/` or `python tests/test_sample_embedding_equivalence.py`."""
 import os
+import subprocess
+import sys
 import tempfile
 import warnings
 
@@ -93,6 +95,35 @@ def test_soft_composition_bit_identical_odd_and_even():
             assert np.array_equal(old, soft_composition(Z, C, rows, n_threads=threads)), (n, K, threads)
 
 
+def test_kmeans_centers_pair_matches_serial_fits():
+    from sampledisco.sample_embedding.blocks import kmeans_centers, kmeans_centers_pair
+    Z = np.random.default_rng(3).normal(size=(5000, 8)).astype(np.float32)
+    pair = kmeans_centers_pair(Z, 6, 12, 0, n_threads=4)
+    assert np.array_equal(pair[0], kmeans_centers(Z, 6, 0)) and np.array_equal(pair[1], kmeans_centers(Z, 12, 1))
+
+
+_NO_WORKER_IMPORTS = """
+import sys, threading
+import numpy as np
+from sampledisco.sample_embedding.blocks import kmeans_centers_pair
+assert "sklearn.cluster" not in sys.modules
+seen = []
+sys.addaudithook(lambda ev, args: ev == "import" and seen.append((args[0], threading.current_thread())))
+kmeans_centers_pair(np.random.default_rng(3).normal(size=(5000, 8)).astype(np.float32), 6, 12, 0, n_threads=4)
+bad = sorted({m for m, t in seen if t is not threading.main_thread()})
+assert not bad, f"{len(bad)} modules first imported in worker threads, e.g. {bad[:5]}"
+"""
+
+
+def test_kmeans_centers_pair_no_imports_in_workers():
+    """A first import inside a worker thread can deadlock against a threadpoolctl scan in the other (0.4.0)."""
+    import sampledisco
+    env = dict(os.environ, PYTHONPATH=os.pathsep.join(
+        [os.path.dirname(os.path.dirname(sampledisco.__file__)), os.environ.get("PYTHONPATH", "")]))
+    r = subprocess.run([sys.executable, "-c", _NO_WORKER_IMPORTS], env=env, capture_output=True, text=True, timeout=300)
+    assert r.returncode == 0, r.stderr[-1500:]
+
+
 def _write_old_cell_h5ad(out):
     a = make_adata()
     a.uns["X_DR_sample"] = pd.DataFrame({"PC1": [1.0]}, index=["old"])