wavedl 1.4.1__py3-none-any.whl → 1.4.3__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__ = "1.4.1"
+__version__ = "1.4.3"
 __author__ = "Ductho Le"
 __email__ = "ductho.le@outlook.com"
 

wavedl/hpo.py

@@ -145,6 +145,7 @@ def create_objective(args):
         # Use temporary directory for trial output
         with tempfile.TemporaryDirectory() as tmpdir:
             cmd.extend(["--output_dir", tmpdir])
+            history_file = Path(tmpdir) / "training_history.csv"
 
             # Run training
             try:
@@ -156,29 +157,55 @@ def create_objective(args):
                     cwd=Path(__file__).parent,
                 )
 
-                # Parse validation loss from output
-                # Look for "Best val_loss: X.XXXX" in stdout
+                # Read best val_loss from training_history.csv (reliable machine-readable)
                 val_loss = None
-                for line in result.stdout.split("\n"):
-                    if "Best val_loss:" in line:
-                        try:
-                            val_loss = float(line.split(":")[-1].strip())
-                        except ValueError:
-                            pass
-                    # Also check for final validation loss
-                    if "val_loss=" in line.lower():
-                        try:
-                            # Extract number after val_loss=
-                            parts = line.lower().split("val_loss=")
-                            if len(parts) > 1:
-                                val_str = parts[1].split()[0].strip(",")
-                                val_loss = float(val_str)
-                        except (ValueError, IndexError):
-                            pass
+                if history_file.exists():
+                    try:
+                        import csv
+
+                        with open(history_file) as f:
+                            reader = csv.DictReader(f)
+                            val_losses = []
+                            for row in reader:
+                                if "val_loss" in row:
+                                    try:
+                                        val_losses.append(float(row["val_loss"]))
+                                    except (ValueError, TypeError):
+                                        pass
+                            if val_losses:
+                                val_loss = min(val_losses)  # Best (minimum) val_loss
+                    except Exception as e:
+                        print(f"Trial {trial.number}: Error reading history: {e}")
+
+                if val_loss is None:
+                    # Fallback: parse stdout for training log format
+                    # Pattern: "epoch | train_loss | val_loss | ..."
+                    # Use regex to avoid false positives from unrelated lines
+                    import re
+
+                    # Match lines like: "  42  | 0.0123   | 0.0156   | ..."
+                    log_pattern = re.compile(
+                        r"^\s*\d+\s*\|\s*[\d.]+\s*\|\s*([\d.]+)\s*\|"
+                    )
+                    val_losses_stdout = []
+                    for line in result.stdout.split("\n"):
+                        match = log_pattern.match(line)
+                        if match:
+                            try:
+                                val_losses_stdout.append(float(match.group(1)))
+                            except ValueError:
+                                continue
+                    if val_losses_stdout:
+                        val_loss = min(val_losses_stdout)
 
                 if val_loss is None:
                     # Training failed or no loss found
-                    print(f"Trial {trial.number}: Training failed")
+                    print(f"Trial {trial.number}: Training failed (no val_loss found)")
+                    if result.returncode != 0:
+                        # Show last few lines of stderr for debugging
+                        stderr_lines = result.stderr.strip().split("\n")[-3:]
+                        for line in stderr_lines:
+                            print(f"  stderr: {line}")
                     return float("inf")
 
                 print(f"Trial {trial.number}: val_loss={val_loss:.6f}")

wavedl/train.py

@@ -37,9 +37,54 @@ Author: Ductho Le (ductho.le@outlook.com)
 
 from __future__ import annotations
 
+# =============================================================================
+# HPC Environment Setup (MUST be before any library imports)
+# =============================================================================
+# Set writable cache directories for matplotlib and fontconfig ONLY when
+# the default paths are not writable (common on HPC clusters).
+import os
+import tempfile
+
+
+def _setup_cache_dir(env_var: str, default_subpath: str) -> None:
+    """Set cache directory only if default path is not writable."""
+    if env_var in os.environ:
+        return  # User already set, respect their choice
+
+    # Check if default home config path is writable
+    home = os.path.expanduser("~")
+    default_path = os.path.join(home, ".config", default_subpath)
+    default_parent = os.path.dirname(default_path)
+
+    # If default path or its parent is writable, let the library use defaults
+    if (
+        os.access(default_path, os.W_OK)
+        or (os.path.exists(default_parent) and os.access(default_parent, os.W_OK))
+        or os.access(os.path.join(home, ".config"), os.W_OK)
+    ):
+        return
+
+    # Default not writable - find alternative location
+    for cache_base in [
+        os.environ.get("SCRATCH"),
+        os.environ.get("SLURM_TMPDIR"),
+        tempfile.gettempdir(),
+    ]:
+        if cache_base and os.access(cache_base, os.W_OK):
+            cache_path = os.path.join(cache_base, f".{default_subpath}")
+            os.makedirs(cache_path, exist_ok=True)
+            os.environ[env_var] = cache_path
+            return
+
+
+_setup_cache_dir("MPLCONFIGDIR", "matplotlib")
+_setup_cache_dir("FONTCONFIG_CACHE", "fontconfig")
+
+# =============================================================================
+# Standard imports (after environment setup)
+# =============================================================================
 import argparse
 import logging
-import os
 import pickle
 import shutil
 import sys
@@ -47,6 +92,10 @@ import time
 import warnings
 from typing import Any
 
+
+# Suppress Pydantic warnings from accelerate's internal Field() usage
+warnings.filterwarnings("ignore", category=UserWarning, module="pydantic")
+
 import matplotlib.pyplot as plt
 import numpy as np
 import pandas as pd
@@ -582,9 +631,9 @@ def main():
     # Torch 2.0 compilation (requires compatible Triton on GPU)
     if args.compile:
         try:
-            # Test if Triton is available AND compatible with this PyTorch version
-            # PyTorch needs triton_key from triton.compiler.compiler
-            from triton.compiler.compiler import triton_key
+            # Test if Triton is available - just import the package
+            # Different Triton versions have different internal APIs, so just check base import
+            import triton
 
             model = torch.compile(model)
             if accelerator.is_main_process:
@@ -851,7 +900,7 @@ def main():
             val_mae_sum = torch.zeros(out_dim, device=accelerator.device)
             val_samples = 0
 
-            # Accumulate predictions locally, gather ONCE at end (reduces sync overhead)
+            # Accumulate predictions locally ON CPU to prevent GPU OOM
             local_preds = []
             local_targets = []
 
@@ -867,17 +916,23 @@ def main():
                     mae_batch = torch.abs((pred - y) * phys_scale).sum(dim=0)
                     val_mae_sum += mae_batch
 
-                    # Store locally (no GPU sync per batch)
-                    local_preds.append(pred)
-                    local_targets.append(y)
+                    # Store on CPU (critical for large val sets)
+                    local_preds.append(pred.detach().cpu())
+                    local_targets.append(y.detach().cpu())
+
+            # Concatenate locally on CPU (no GPU memory spike)
+            cpu_preds = torch.cat(local_preds)
+            cpu_targets = torch.cat(local_targets)
 
-            # Single gather at end of validation (2 syncs instead of 2×num_batches)
-            all_local_preds = torch.cat(local_preds)
-            all_local_targets = torch.cat(local_targets)
-            all_preds = accelerator.gather_for_metrics(all_local_preds)
-            all_targets = accelerator.gather_for_metrics(all_local_targets)
+            # Gather predictions and targets across all ranks
+            # Use accelerator.gather (works with all accelerate versions)
+            gpu_preds = cpu_preds.to(accelerator.device)
+            gpu_targets = cpu_targets.to(accelerator.device)
+            all_preds_gathered = accelerator.gather(gpu_preds).cpu()
+            all_targets_gathered = accelerator.gather(gpu_targets).cpu()
+            gathered = [(all_preds_gathered, all_targets_gathered)]
 
-            # Synchronize validation metrics
+            # Synchronize validation metrics (scalars only - efficient)
             val_loss_scalar = val_loss_sum.item()
             val_metrics = torch.cat(
                 [
@@ -900,9 +955,14 @@ def main():
 
             # ==================== LOGGING & CHECKPOINTING ====================
             if accelerator.is_main_process:
-                # Scientific metrics - cast to float32 before numpy (bf16 can't convert)
-                y_pred = all_preds.float().cpu().numpy()
-                y_true = all_targets.float().cpu().numpy()
+                # Concatenate gathered tensors from all ranks (only on rank 0)
+                # gathered is list of tuples: [(preds_rank0, targs_rank0), (preds_rank1, targs_rank1), ...]
+                all_preds = torch.cat([item[0] for item in gathered])
+                all_targets = torch.cat([item[1] for item in gathered])
+
+                # Scientific metrics - cast to float32 before numpy
+                y_pred = all_preds.float().numpy()
+                y_true = all_targets.float().numpy()
 
                 # Trim DDP padding
                 real_len = len(val_dl.dataset)

wavedl/utils/data.py

@@ -735,7 +735,7 @@ def load_test_data(
     try:
         inp, outp = source.load(path)
     except KeyError:
-        # Try with just inputs if outputs not found
+        # Try with just inputs if outputs not found (inference-only mode)
         if format == "npz":
             data = np.load(path, allow_pickle=True)
             keys = list(data.keys())
@@ -751,6 +751,54 @@ def load_test_data(
                 )
             out_key = DataSource._find_key(keys, custom_output_keys)
             outp = data[out_key] if out_key else None
+        elif format == "hdf5":
+            # HDF5: input-only loading for inference
+            with h5py.File(path, "r") as f:
+                keys = list(f.keys())
+                inp_key = DataSource._find_key(keys, custom_input_keys)
+                if inp_key is None:
+                    raise KeyError(
+                        f"Input key not found. Tried: {custom_input_keys}. Found: {keys}"
+                    )
+                # Check size - load_test_data is eager, large files should use DataLoader
+                n_samples = f[inp_key].shape[0]
+                if n_samples > 100000:
+                    raise ValueError(
+                        f"Dataset has {n_samples:,} samples. load_test_data() loads "
+                        f"everything into RAM which may cause OOM. For large inference "
+                        f"sets, use a DataLoader with HDF5Source.load_mmap() instead."
+                    )
+                inp = f[inp_key][:]
+                out_key = DataSource._find_key(keys, custom_output_keys)
+                outp = f[out_key][:] if out_key else None
+        elif format == "mat":
+            # MAT v7.3: input-only loading with proper sparse handling
+            mat_source = MATSource()
+            with h5py.File(path, "r") as f:
+                keys = list(f.keys())
+                inp_key = DataSource._find_key(keys, custom_input_keys)
+                if inp_key is None:
+                    raise KeyError(
+                        f"Input key not found. Tried: {custom_input_keys}. Found: {keys}"
+                    )
+                # Check size - load_test_data is eager, large files should use DataLoader
+                n_samples = f[inp_key].shape[-1]  # MAT is transposed
+                if n_samples > 100000:
+                    raise ValueError(
+                        f"Dataset has {n_samples:,} samples. load_test_data() loads "
+                        f"everything into RAM which may cause OOM. For large inference "
+                        f"sets, use a DataLoader with MATSource.load_mmap() instead."
+                    )
+                # Use _load_dataset for sparse support and proper transpose
+                inp = mat_source._load_dataset(f, inp_key)
+                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
+                else:
+                    outp = None
         else:
             raise
 
@@ -949,6 +997,15 @@ def prepare_data(
             with open(META_FILE, "rb") as f:
                 meta = pickle.load(f)
             cached_data_path = meta.get("data_path", None)
+            cached_file_size = meta.get("file_size", None)
+            cached_file_mtime = meta.get("file_mtime", None)
+
+            # Get current file stats
+            current_stats = os.stat(args.data_path)
+            current_size = current_stats.st_size
+            current_mtime = current_stats.st_mtime
+
+            # Check if data path changed
             if cached_data_path != os.path.abspath(args.data_path):
                 if accelerator.is_main_process:
                     logger.warning(
@@ -958,6 +1015,23 @@ def prepare_data(
                         f"   Invalidating cache and regenerating..."
                     )
                 cache_exists = False
+            # Check if file was modified (size or mtime changed)
+            elif cached_file_size is not None and cached_file_size != current_size:
+                if accelerator.is_main_process:
+                    logger.warning(
+                        f"⚠️  Data file size changed!\n"
+                        f"   Cached size: {cached_file_size:,} bytes\n"
+                        f"   Current size: {current_size:,} bytes\n"
+                        f"   Invalidating cache and regenerating..."
+                    )
+                cache_exists = False
+            elif cached_file_mtime is not None and cached_file_mtime != current_mtime:
+                if accelerator.is_main_process:
+                    logger.warning(
+                        "⚠️  Data file was modified!\n"
+                        "   Cache may be stale, regenerating..."
+                    )
+                cache_exists = False
         except Exception:
             cache_exists = False
 
@@ -1053,13 +1127,16 @@ def prepare_data(
                 f"   Shape Detected: {full_shape} [{dim_type}] | Output Dim: {out_dim}"
             )
 
-            # Save metadata (including data path for cache validation)
+            # Save metadata (including data path, size, mtime for cache validation)
+            file_stats = os.stat(args.data_path)
             with open(META_FILE, "wb") as f:
                 pickle.dump(
                     {
                         "shape": full_shape,
                         "out_dim": out_dim,
                         "data_path": os.path.abspath(args.data_path),
+                        "file_size": file_stats.st_size,
+                        "file_mtime": file_stats.st_mtime,
                     },
                     f,
                 )

{wavedl-1.4.1.dist-info → wavedl-1.4.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: wavedl
-Version: 1.4.1
+Version: 1.4.3
 Summary: A Scalable Deep Learning Framework for Wave-Based Inverse Problems
 Author: Ductho Le
 License: MIT
@@ -57,6 +57,7 @@ Requires-Dist: triton>=2.0.0; sys_platform == "linux"
 [![Lint](https://img.shields.io/github/actions/workflow/status/ductho-le/WaveDL/lint.yml?branch=main&style=plastic&logo=ruff&logoColor=white&label=Lint)](https://github.com/ductho-le/WaveDL/actions/workflows/lint.yml)
 [![Try it on Colab](https://img.shields.io/badge/Try_it_on_Colab-8E44AD?style=plastic&logo=googlecolab&logoColor=white)](https://colab.research.google.com/github/ductho-le/WaveDL/blob/main/notebooks/demo.ipynb)
 <br>
+[![Downloads](https://img.shields.io/pepy/dt/wavedl?style=plastic&logo=pypi&logoColor=white&color=9ACD32)](https://pepy.tech/project/wavedl)
 [![License: MIT](https://img.shields.io/badge/License-MIT-orange.svg?style=plastic)](LICENSE)
 [![DOI](https://img.shields.io/badge/DOI-10.5281/zenodo.18012338-008080.svg?style=plastic)](https://doi.org/10.5281/zenodo.18012338)
 

{wavedl-1.4.1.dist-info → wavedl-1.4.3.dist-info}/RECORD

@@ -1,8 +1,8 @@
-wavedl/__init__.py,sha256=2LU5rtHKoYgXBAZ4zGNtFcHjrTtmmYskXnaURHEwkNc,1177
+wavedl/__init__.py,sha256=7uvgr9r21qLGu6RIZDUfQBhg1vXNAfmIWl_P6BMj_KQ,1177
 wavedl/hpc.py,sha256=de_GKERX8GS10sXRX9yXiGzMnk1jjq8JPzRw7QDs6d4,7967
-wavedl/hpo.py,sha256=aZoa_Oto_anZpIhz-YM6kN8KxQXTolUvDEyg3NXwBrY,11542
+wavedl/hpo.py,sha256=YJXsnSGEBSVUqp_2ah7zu3_VClAUqZrdkuzDaSqQUjU,12952
 wavedl/test.py,sha256=jZmRJaivYYTMMTaccCi0yQjHOfp0a9YWR1wAPeKFH-k,36246
-wavedl/train.py,sha256=e0tX7_j2gkuYpPjZJqGoDV8arAe4bc4YVRMyrg-RcRY,46402
+wavedl/train.py,sha256=TjLABBPCqu9r7FEWlxJlKsT7uAMo6hiDxRfii2SKXe4,49052
 wavedl/models/__init__.py,sha256=lfSohEnAUztO14nuwayMJhPjpgySzRN3jGiyAUuBmAU,3206
 wavedl/models/_template.py,sha256=J_D8taSPmV8lBaucN_vU-WiG98iFr7CJrZVNNX_Tdts,4600
 wavedl/models/base.py,sha256=T9iDF9IQM2MYucG_ggQd31rieUkB2fob-nkHyNIl2ak,7337
@@ -23,15 +23,15 @@ wavedl/models/vit.py,sha256=0C3GZk11VsYFTl14d86Wtl1Zk1T5rYJjvkaEfEN4N3k,11100
 wavedl/utils/__init__.py,sha256=YMgzuwndjr64kt9k0_6_9PMJYTVdiaH5veSMff_ZycA,3051
 wavedl/utils/config.py,sha256=fMoucikIQHn85mVhGMa7TnXTuFDcEEPjfXk2EjbkJR0,10591
 wavedl/utils/cross_validation.py,sha256=117ac9KDzaIaqhtP8ZRs15Xpqmq5fLpX2-vqkNvtMaU,17487
-wavedl/utils/data.py,sha256=9LrB9MC6jRZzbRSc9xiGzJWoh8FahwP_68REqBAT3Os,44131
+wavedl/utils/data.py,sha256=_OaWvU5oFVJW0NwM5WyDD0Kb1hy5MgvJIFpzvJGux9w,48214
 wavedl/utils/distributed.py,sha256=7wQ3mRjkp_xjPSxDWMnBf5dSkAGUaTzntxbz0BhC5v0,4145
 wavedl/utils/losses.py,sha256=5762M-TBC_hz6uyj1NPbU1vZeFOJQq7fR3-j7OygJRo,7254
 wavedl/utils/metrics.py,sha256=mkCpqZwl_XUpNvA5Ekjf7y-HqApafR7eR6EuA8cBdM8,37287
 wavedl/utils/optimizers.py,sha256=PyIkJ_hRhFi_Fio81Gy5YQNhcME0JUUEl8OTSyu-0RA,6323
 wavedl/utils/schedulers.py,sha256=e6Sf0yj8VOqkdwkUHLMyUfGfHKTX4NMr-zfgxWqCTYI,7659
-wavedl-1.4.1.dist-info/LICENSE,sha256=cEUCvcvH-9BT9Y-CNGY__PwWONCKu9zsoIqWA-NeHJ4,1066
-wavedl-1.4.1.dist-info/METADATA,sha256=FEafy9hY2su6bB8iS8VNZceLpTs9E7nhVaejsOEHTUM,40245
-wavedl-1.4.1.dist-info/WHEEL,sha256=beeZ86-EfXScwlR_HKu4SllMC9wUEj_8Z_4FJ3egI2w,91
-wavedl-1.4.1.dist-info/entry_points.txt,sha256=f1RNDkXFZwBzrBzTMFocJ6xhfTvTmaEDTi5YyDEUaF8,140
-wavedl-1.4.1.dist-info/top_level.txt,sha256=ccneUt3D5Qzbh3bsBSSrq9bqrhGiogcWKY24ZC4Q6Xw,7
-wavedl-1.4.1.dist-info/RECORD,,
+wavedl-1.4.3.dist-info/LICENSE,sha256=cEUCvcvH-9BT9Y-CNGY__PwWONCKu9zsoIqWA-NeHJ4,1066
+wavedl-1.4.3.dist-info/METADATA,sha256=vs3nt8R2O5lD7q-si9M5ChyrriIKm0fFzDwi4HVIYxw,40386
+wavedl-1.4.3.dist-info/WHEEL,sha256=beeZ86-EfXScwlR_HKu4SllMC9wUEj_8Z_4FJ3egI2w,91
+wavedl-1.4.3.dist-info/entry_points.txt,sha256=f1RNDkXFZwBzrBzTMFocJ6xhfTvTmaEDTi5YyDEUaF8,140
+wavedl-1.4.3.dist-info/top_level.txt,sha256=ccneUt3D5Qzbh3bsBSSrq9bqrhGiogcWKY24ZC4Q6Xw,7
+wavedl-1.4.3.dist-info/RECORD,,