wavedl 1.5.4__py3-none-any.whl → 1.5.5__py3-none-any.whl

wavedl/__init__.py

@@ -18,7 +18,7 @@ For inference:
     # or: python -m wavedl.test --checkpoint best_checkpoint --data_path test.npz
 """
 
-__version__ = "1.5.4"
+__version__ = "1.5.5"
 __author__ = "Ductho Le"
 __email__ = "ductho.le@outlook.com"
 

wavedl/hpo.py

@@ -175,13 +175,14 @@ def create_objective(args):
                 env["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
 
             # Run training
+            # Note: We inherit the user's cwd instead of setting cwd=Path(__file__).parent
+            # because site-packages may be read-only and train.py creates cache directories
             try:
                 result = subprocess.run(
                     cmd,
                     capture_output=True,
                     text=True,
                     timeout=args.timeout,
-                    cwd=Path(__file__).parent,
                     env=env,
                 )
 

wavedl/utils/data.py

@@ -207,6 +207,10 @@ class NPZSource(DataSource):
 
         The error for object arrays happens at ACCESS time, not load time.
         So we need to probe the keys to detect if pickle is required.
+
+        WARNING: When mmap_mode is not None, the returned NpzFile must be kept
+        open for arrays to remain valid. Caller is responsible for closing.
+        For non-mmap loading, use _load_and_copy() instead to avoid leaks.
         """
         data = np.load(path, allow_pickle=False, mmap_mode=mmap_mode)
         try:
@@ -222,6 +226,26 @@ class NPZSource(DataSource):
                 return np.load(path, allow_pickle=True, mmap_mode=mmap_mode)
             raise
 
+    @staticmethod
+    def _load_and_copy(path: str, keys: list[str]) -> dict[str, np.ndarray]:
+        """Load NPZ and copy arrays, ensuring file is properly closed.
+
+        This prevents file descriptor leaks by copying arrays before closing.
+        Use this for eager loading; use _safe_load for memory-mapped access.
+        """
+        data = NPZSource._safe_load(path, keys, mmap_mode=None)
+        try:
+            result = {}
+            for key in keys:
+                if key in data:
+                    arr = data[key]
+                    # Copy ensures we don't hold reference to mmap
+                    result[key] = arr.copy() if hasattr(arr, "copy") else arr
+            return result
+        finally:
+            if hasattr(data, "close"):
+                data.close()
+
     def load(self, path: str) -> tuple[np.ndarray, np.ndarray]:
         """Load NPZ file (pickle enabled only for sparse matrices)."""
         # First pass to find keys without loading data
@@ -238,7 +262,7 @@ class NPZSource(DataSource):
                 f"Found: {keys}"
             )
 
-        data = self._safe_load(path, [input_key, output_key])
+        data = self._load_and_copy(path, [input_key, output_key])
         inp = data[input_key]
         outp = data[output_key]
 
@@ -290,7 +314,7 @@ class NPZSource(DataSource):
                 f"Supported keys: {OUTPUT_KEYS}. Found: {keys}"
             )
 
-        data = self._safe_load(path, [output_key])
+        data = self._load_and_copy(path, [output_key])
         return data[output_key]
 
 
@@ -527,9 +551,17 @@ class MATSource(DataSource):
                 inp = self._load_dataset(f, input_key)
                 outp = self._load_dataset(f, output_key)
 
-                # Handle 1D outputs that become (1, N) after transpose
-                if outp.ndim == 2 and outp.shape[0] == 1:
-                    outp = outp.T
+                # Handle transposed outputs from MATLAB.
+                # Case 1: (1, N) - N samples with 1 target → transpose to (N, 1)
+                # Case 2: (T, 1) - 1 sample with T targets → transpose to (1, T)
+                num_samples = inp.shape[0]  # inp is already transposed
+                if outp.ndim == 2:
+                    if outp.shape[0] == 1 and outp.shape[1] == num_samples:
+                        # 1D vector: (1, N) → (N, 1)
+                        outp = outp.T
+                    elif outp.shape[1] == 1 and outp.shape[0] != num_samples:
+                        # Single sample with multiple targets: (T, 1) → (1, T)
+                        outp = outp.T
 
         except OSError as e:
             raise ValueError(
@@ -614,7 +646,10 @@ class MATSource(DataSource):
                 # Load with sparse matrix support
                 outp = self._load_dataset(f, output_key)
 
-                # Handle 1D outputs
+                # Handle 1D outputs that become (1, N) after transpose.
+                # Note: This method has no input to compare against, so we can't
+                # distinguish single-sample outputs. This is acceptable for training
+                # data where single-sample is unlikely. For inference, use load_test_data.
                 if outp.ndim == 2 and outp.shape[0] == 1:
                     outp = outp.T
 
@@ -775,7 +810,7 @@ def load_test_data(
                 raise KeyError(
                     f"Input key not found. Tried: {custom_input_keys}. Found: {keys}"
                 )
-            data = NPZSource._safe_load(
+            data = NPZSource._load_and_copy(
                 path, [inp_key] + ([out_key] if out_key else [])
             )
             inp = data[inp_key]
@@ -824,8 +859,17 @@ def load_test_data(
                 inp = mat_source._load_dataset(f, inp_key)
                 if out_key:
                     outp = mat_source._load_dataset(f, out_key)
-                    if outp.ndim == 2 and outp.shape[0] == 1:
-                        outp = outp.T
+                    # Handle transposed outputs from MATLAB
+                    # Case 1: (1, N) - N samples with 1 target → transpose to (N, 1)
+                    # Case 2: (T, 1) - 1 sample with T targets → transpose to (1, T)
+                    num_samples = inp.shape[0]
+                    if outp.ndim == 2:
+                        if outp.shape[0] == 1 and outp.shape[1] == num_samples:
+                            # 1D vector: (1, N) → (N, 1)
+                            outp = outp.T
+                        elif outp.shape[1] == 1 and outp.shape[0] != num_samples:
+                            # Single sample with multiple targets: (T, 1) → (1, T)
+                            outp = outp.T
                 else:
                     outp = None
         else:
@@ -844,7 +888,7 @@ def load_test_data(
                 )
             out_key = DataSource._find_key(keys, custom_output_keys)
             keys_to_probe = [inp_key] + ([out_key] if out_key else [])
-            data = NPZSource._safe_load(path, keys_to_probe)
+            data = NPZSource._load_and_copy(path, keys_to_probe)
             inp = data[inp_key]
             if inp.dtype == object:
                 inp = np.array(
@@ -894,9 +938,17 @@ def load_test_data(
                 out_key = DataSource._find_key(keys, custom_output_keys)
                 if out_key:
                     outp = mat_source._load_dataset(f, out_key)
-                    # Handle 1D outputs that become (1, N) after transpose
-                    if outp.ndim == 2 and outp.shape[0] == 1:
-                        outp = outp.T
+                    # Handle transposed outputs from MATLAB
+                    # Case 1: (1, N) - N samples with 1 target → transpose to (N, 1)
+                    # Case 2: (T, 1) - 1 sample with T targets → transpose to (1, T)
+                    num_samples = inp.shape[0]
+                    if outp.ndim == 2:
+                        if outp.shape[0] == 1 and outp.shape[1] == num_samples:
+                            # 1D vector: (1, N) → (N, 1)
+                            outp = outp.T
+                        elif outp.shape[1] == 1 and outp.shape[0] != num_samples:
+                            # Single sample with multiple targets: (T, 1) → (1, T)
+                            outp = outp.T
                 else:
                     outp = None
         else:
@@ -1153,6 +1205,18 @@ def prepare_data(
                         logger.warning(
                             f"   Failed to remove stale cache {stale_file}: {e}"
                         )
+
+            # Fail explicitly if stale cache files couldn't be removed
+            # This prevents silent reuse of outdated data
+            remaining_stale = [
+                f for f in [CACHE_FILE, SCALER_FILE] if os.path.exists(f)
+            ]
+            if remaining_stale:
+                raise RuntimeError(
+                    f"Cannot regenerate cache: stale files could not be removed. "
+                    f"Please manually delete: {remaining_stale}"
+                )
+
             # RANK 0: Create cache (can take a long time for large datasets)
             # Other ranks will wait at the barrier below
 

wavedl/utils/metrics.py

@@ -815,7 +815,28 @@ def plot_qq(
 
         # Standardize errors for QQ plot
         err = errors[:, i]
-        standardized = (err - np.mean(err)) / np.std(err)
+        std_err = np.std(err)
+
+        # Guard against zero variance (constant errors)
+        if std_err < 1e-10:
+            title = (
+                param_names[i] if param_names and i < len(param_names) else f"Param {i}"
+            )
+            ax.text(
+                0.5,
+                0.5,
+                "Zero variance\n(constant errors)",
+                ha="center",
+                va="center",
+                fontsize=10,
+                transform=ax.transAxes,
+            )
+            ax.set_title(f"{title}\n(zero variance)")
+            ax.set_xlabel("Theoretical Quantiles")
+            ax.set_ylabel("Sample Quantiles")
+            continue
+
+        standardized = (err - np.mean(err)) / std_err
 
         # Calculate theoretical quantiles and sample quantiles
         (osm, osr), (slope, intercept, r) = stats.probplot(standardized, dist="norm")

{wavedl-1.5.4.dist-info → wavedl-1.5.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: wavedl
-Version: 1.5.4
+Version: 1.5.5
 Summary: A Scalable Deep Learning Framework for Wave-Based Inverse Problems
 Author: Ductho Le
 License: MIT
@@ -388,7 +388,7 @@ WaveDL/
 ├── configs/                      # YAML config templates
 ├── examples/                     # Ready-to-run examples
 ├── notebooks/                    # Jupyter notebooks
-├── unit_tests/                   # Pytest test suite (725 tests)
+├── unit_tests/                   # Pytest test suite (731 tests)
 │
 ├── pyproject.toml                # Package config, dependencies
 ├── CHANGELOG.md                  # Version history

{wavedl-1.5.4.dist-info → wavedl-1.5.5.dist-info}/RECORD

@@ -1,6 +1,6 @@
-wavedl/__init__.py,sha256=L3ckuWk3BDr6h9oiADkGP_JKcGSF669qDkuzofh86IU,1177
+wavedl/__init__.py,sha256=RTePiYlzCrUofbGSYWAAqoKeeyYjqEPzuXyze6ai324,1177
 wavedl/hpc.py,sha256=6rV38nozzMt0-jKZbVJNwvQZXK0wUsIZmr9lgWN_XUw,9212
-wavedl/hpo.py,sha256=DGCGyt2yhr3WAifAuljhE26gg07CHdaQW4wpDaTKbyo,14968
+wavedl/hpo.py,sha256=CZF0MZwTGMOrPGDveUXZFbGHwLHj1FcJTCBKVVEtLWg,15105
 wavedl/test.py,sha256=WIHG3HWT-uF399FQApPpxjggBVFn59cC54HAL4990QU,38550
 wavedl/train.py,sha256=7AVaCORFUv2_IgdYSPKdHLxbi11GzMOyu4RcNc4Uf_I,55963
 wavedl/models/__init__.py,sha256=lfSohEnAUztO14nuwayMJhPjpgySzRN3jGiyAUuBmAU,3206
@@ -24,15 +24,15 @@ wavedl/utils/__init__.py,sha256=s5R9bRmJ8GNcJrD3OSAOXzwZJIXZbdYrAkZnus11sVQ,3300
 wavedl/utils/config.py,sha256=AsGwb3XtxmbTLb59BLl5AA4wzMNgVTpl7urOJ6IGqfM,10901
 wavedl/utils/constraints.py,sha256=Pof5hzeTSGsPY_E6Sc8iMQDaXc_zfEasQI2tCszk_gw,17614
 wavedl/utils/cross_validation.py,sha256=gwXSFTx5oxWndPjWLJAJzB6nnq2f1t9f86SbjbF-jNI,18475
-wavedl/utils/data.py,sha256=JusSrIZd98t9oiN0xTy2V2mfVyuBCIu0MLAQGcaC0vQ,54194
+wavedl/utils/data.py,sha256=cmJ6tUw4Tcxj-l3Xsphs1Dnlx1MzxOPvk8etD5KXFNs,57686
 wavedl/utils/distributed.py,sha256=7wQ3mRjkp_xjPSxDWMnBf5dSkAGUaTzntxbz0BhC5v0,4145
 wavedl/utils/losses.py,sha256=5762M-TBC_hz6uyj1NPbU1vZeFOJQq7fR3-j7OygJRo,7254
-wavedl/utils/metrics.py,sha256=EJmJvF7gACQsUoKYldlladN_SbnRiuE-Smj0eSnbraQ,39394
+wavedl/utils/metrics.py,sha256=El2NYsulH5jxBhC1gCAMcS8C-yxEjuSC930LhsKYQrY,40059
 wavedl/utils/optimizers.py,sha256=PyIkJ_hRhFi_Fio81Gy5YQNhcME0JUUEl8OTSyu-0RA,6323
 wavedl/utils/schedulers.py,sha256=e6Sf0yj8VOqkdwkUHLMyUfGfHKTX4NMr-zfgxWqCTYI,7659
-wavedl-1.5.4.dist-info/LICENSE,sha256=cEUCvcvH-9BT9Y-CNGY__PwWONCKu9zsoIqWA-NeHJ4,1066
-wavedl-1.5.4.dist-info/METADATA,sha256=D7_MbjGWyVEIEH2m23GrJInZO4pcfHAINlY1FIUgD-A,45604
-wavedl-1.5.4.dist-info/WHEEL,sha256=beeZ86-EfXScwlR_HKu4SllMC9wUEj_8Z_4FJ3egI2w,91
-wavedl-1.5.4.dist-info/entry_points.txt,sha256=f1RNDkXFZwBzrBzTMFocJ6xhfTvTmaEDTi5YyDEUaF8,140
-wavedl-1.5.4.dist-info/top_level.txt,sha256=ccneUt3D5Qzbh3bsBSSrq9bqrhGiogcWKY24ZC4Q6Xw,7
-wavedl-1.5.4.dist-info/RECORD,,
+wavedl-1.5.5.dist-info/LICENSE,sha256=cEUCvcvH-9BT9Y-CNGY__PwWONCKu9zsoIqWA-NeHJ4,1066
+wavedl-1.5.5.dist-info/METADATA,sha256=0e7E8zLd-GlcR5Hbgp5VDYGGr36_9NzKBTShsG4xuQs,45604
+wavedl-1.5.5.dist-info/WHEEL,sha256=beeZ86-EfXScwlR_HKu4SllMC9wUEj_8Z_4FJ3egI2w,91
+wavedl-1.5.5.dist-info/entry_points.txt,sha256=f1RNDkXFZwBzrBzTMFocJ6xhfTvTmaEDTi5YyDEUaF8,140
+wavedl-1.5.5.dist-info/top_level.txt,sha256=ccneUt3D5Qzbh3bsBSSrq9bqrhGiogcWKY24ZC4Q6Xw,7
+wavedl-1.5.5.dist-info/RECORD,,