wavedl 1.5.1__tar.gz → 1.5.3__tar.gz

{wavedl-1.5.1/src/wavedl.egg-info → wavedl-1.5.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: wavedl
-Version: 1.5.1
+Version: 1.5.3
 Summary: A Scalable Deep Learning Framework for Wave-Based Inverse Problems
 Author: Ductho Le
 License: MIT
@@ -301,8 +301,8 @@ python -m wavedl.test --checkpoint <checkpoint_folder> --data_path <test_data> \
 
 **Requirements** (your model must):
 1. Inherit from `BaseModel`
-2. Accept `in_channels`, `num_outputs`, `input_shape` in `__init__`
-3. Return a tensor of shape `(batch, num_outputs)` from `forward()`
+2. Accept `in_shape`, `out_size` in `__init__`
+3. Return a tensor of shape `(batch, out_size)` from `forward()`
 
 ---
 
@@ -315,23 +315,22 @@ from wavedl.models import BaseModel, register_model
 
 @register_model("my_model")  # This name is used with --model flag
 class MyModel(BaseModel):
-    def __init__(self, in_channels, num_outputs, input_shape):
-        # in_channels: number of input channels (auto-detected from data)
-        # num_outputs: number of parameters to predict (auto-detected from data)
-        # input_shape: spatial dimensions, e.g., (128,) or (64, 64) or (32, 32, 32)
-        super().__init__(in_channels, num_outputs, input_shape)
-
-        # Define your layers (this is just an example)
-        self.conv1 = nn.Conv2d(in_channels, 64, 3, padding=1)
+    def __init__(self, in_shape, out_size, **kwargs):
+        # in_shape: spatial dimensions, e.g., (128,) or (64, 64) or (32, 32, 32)
+        # out_size: number of parameters to predict (auto-detected from data)
+        super().__init__(in_shape, out_size)
+
+        # Define your layers (this is just an example for 2D)
+        self.conv1 = nn.Conv2d(1, 64, 3, padding=1)  # Input always has 1 channel
         self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
-        self.fc = nn.Linear(128, num_outputs)
+        self.fc = nn.Linear(128, out_size)
 
     def forward(self, x):
-        # Input x has shape: (batch, in_channels, *input_shape)
+        # Input x has shape: (batch, 1, *in_shape)
         x = F.relu(self.conv1(x))
         x = F.relu(self.conv2(x))
         x = x.mean(dim=[-2, -1])  # Global average pooling
-        return self.fc(x)  # Output shape: (batch, num_outputs)
+        return self.fc(x)  # Output shape: (batch, out_size)
 ```
 
 **Step 2: Train**
@@ -573,14 +572,19 @@ WaveDL automatically enables performance optimizations for modern GPUs:
 </details>
 
 <details>
-<summary><b>Environment Variables (wavedl-hpc)</b></summary>
+<summary><b>HPC CLI Arguments (wavedl-hpc)</b></summary>
+
+| Argument | Default | Description |
+|----------|---------|-------------|
+| `--num_gpus` | **Auto-detected** | Number of GPUs to use. By default, automatically detected via `nvidia-smi`. Set explicitly to override |
+| `--num_machines` | `1` | Number of machines in distributed setup |
+| `--mixed_precision` | `bf16` | Precision mode: `bf16`, `fp16`, or `no` |
+| `--dynamo_backend` | `no` | PyTorch Dynamo backend |
+
+**Environment Variables (for logging):**
 
 | Variable | Default | Description |
 |----------|---------|-------------|
-| `NUM_GPUS` | **Auto-detected** | Number of GPUs to use. By default, automatically detected via `nvidia-smi`. Set explicitly to override (e.g., `NUM_GPUS=2`) |
-| `NUM_MACHINES` | `1` | Number of machines in distributed setup |
-| `MIXED_PRECISION` | `bf16` | Precision mode: `bf16`, `fp16`, or `no` |
-| `DYNAMO_BACKEND` | `no` | PyTorch Dynamo backend |
 | `WANDB_MODE` | `offline` | WandB mode: `offline` or `online` |
 
 </details>

{wavedl-1.5.1 → wavedl-1.5.3}/README.md

@@ -256,8 +256,8 @@ python -m wavedl.test --checkpoint <checkpoint_folder> --data_path <test_data> \
 
 **Requirements** (your model must):
 1. Inherit from `BaseModel`
-2. Accept `in_channels`, `num_outputs`, `input_shape` in `__init__`
-3. Return a tensor of shape `(batch, num_outputs)` from `forward()`
+2. Accept `in_shape`, `out_size` in `__init__`
+3. Return a tensor of shape `(batch, out_size)` from `forward()`
 
 ---
 
@@ -270,23 +270,22 @@ from wavedl.models import BaseModel, register_model
 
 @register_model("my_model")  # This name is used with --model flag
 class MyModel(BaseModel):
-    def __init__(self, in_channels, num_outputs, input_shape):
-        # in_channels: number of input channels (auto-detected from data)
-        # num_outputs: number of parameters to predict (auto-detected from data)
-        # input_shape: spatial dimensions, e.g., (128,) or (64, 64) or (32, 32, 32)
-        super().__init__(in_channels, num_outputs, input_shape)
-
-        # Define your layers (this is just an example)
-        self.conv1 = nn.Conv2d(in_channels, 64, 3, padding=1)
+    def __init__(self, in_shape, out_size, **kwargs):
+        # in_shape: spatial dimensions, e.g., (128,) or (64, 64) or (32, 32, 32)
+        # out_size: number of parameters to predict (auto-detected from data)
+        super().__init__(in_shape, out_size)
+
+        # Define your layers (this is just an example for 2D)
+        self.conv1 = nn.Conv2d(1, 64, 3, padding=1)  # Input always has 1 channel
         self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
-        self.fc = nn.Linear(128, num_outputs)
+        self.fc = nn.Linear(128, out_size)
 
     def forward(self, x):
-        # Input x has shape: (batch, in_channels, *input_shape)
+        # Input x has shape: (batch, 1, *in_shape)
         x = F.relu(self.conv1(x))
         x = F.relu(self.conv2(x))
         x = x.mean(dim=[-2, -1])  # Global average pooling
-        return self.fc(x)  # Output shape: (batch, num_outputs)
+        return self.fc(x)  # Output shape: (batch, out_size)
 ```
 
 **Step 2: Train**
@@ -528,14 +527,19 @@ WaveDL automatically enables performance optimizations for modern GPUs:
 </details>
 
 <details>
-<summary><b>Environment Variables (wavedl-hpc)</b></summary>
+<summary><b>HPC CLI Arguments (wavedl-hpc)</b></summary>
+
+| Argument | Default | Description |
+|----------|---------|-------------|
+| `--num_gpus` | **Auto-detected** | Number of GPUs to use. By default, automatically detected via `nvidia-smi`. Set explicitly to override |
+| `--num_machines` | `1` | Number of machines in distributed setup |
+| `--mixed_precision` | `bf16` | Precision mode: `bf16`, `fp16`, or `no` |
+| `--dynamo_backend` | `no` | PyTorch Dynamo backend |
+
+**Environment Variables (for logging):**
 
 | Variable | Default | Description |
 |----------|---------|-------------|
-| `NUM_GPUS` | **Auto-detected** | Number of GPUs to use. By default, automatically detected via `nvidia-smi`. Set explicitly to override (e.g., `NUM_GPUS=2`) |
-| `NUM_MACHINES` | `1` | Number of machines in distributed setup |
-| `MIXED_PRECISION` | `bf16` | Precision mode: `bf16`, `fp16`, or `no` |
-| `DYNAMO_BACKEND` | `no` | PyTorch Dynamo backend |
 | `WANDB_MODE` | `offline` | WandB mode: `offline` or `online` |
 
 </details>

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/__init__.py

@@ -18,7 +18,7 @@ For inference:
     # or: python -m wavedl.test --checkpoint best_checkpoint --data_path test.npz
 """
 
-__version__ = "1.5.1"
+__version__ = "1.5.3"
 __author__ = "Ductho Le"
 __email__ = "ductho.le@outlook.com"
 

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/hpc.py

@@ -57,30 +57,35 @@ def setup_hpc_environment() -> None:
     """Configure environment variables for HPC systems.
 
     Handles restricted home directories (e.g., Compute Canada) and
-    offline logging configurations.
+    offline logging configurations. Always uses CWD-based TORCH_HOME
+    since compute nodes typically lack internet access.
     """
-    # Check if home is writable
+    # Use CWD for cache base since HPC compute nodes typically lack internet
+    cache_base = os.getcwd()
+
+    # TORCH_HOME always set to CWD - compute nodes need pre-cached weights
+    os.environ.setdefault("TORCH_HOME", f"{cache_base}/.torch_cache")
+    Path(os.environ["TORCH_HOME"]).mkdir(parents=True, exist_ok=True)
+
+    # Triton/Inductor caches - prevents permission errors with --compile
+    # These MUST be set before any torch.compile calls
+    os.environ.setdefault("TRITON_CACHE_DIR", f"{cache_base}/.triton_cache")
+    os.environ.setdefault("TORCHINDUCTOR_CACHE_DIR", f"{cache_base}/.inductor_cache")
+    Path(os.environ["TRITON_CACHE_DIR"]).mkdir(parents=True, exist_ok=True)
+    Path(os.environ["TORCHINDUCTOR_CACHE_DIR"]).mkdir(parents=True, exist_ok=True)
+
+    # Check if home is writable for other caches
     home = os.path.expanduser("~")
     home_writable = os.access(home, os.W_OK)
 
-    # Use SLURM_TMPDIR if available, otherwise CWD for HPC, or system temp
-    if home_writable:
-        # Local machine - let libraries use defaults
-        cache_base = None
-    else:
-        # HPC with restricted home - use CWD for persistent caches
-        cache_base = os.getcwd()
-
-    # Only set environment variables if home is not writable
-    if cache_base:
-        os.environ.setdefault("TORCH_HOME", f"{cache_base}/.torch_cache")
+    # Other caches only if home is not writable
+    if not home_writable:
         os.environ.setdefault("MPLCONFIGDIR", f"{cache_base}/.matplotlib")
         os.environ.setdefault("FONTCONFIG_CACHE", f"{cache_base}/.fontconfig")
         os.environ.setdefault("XDG_CACHE_HOME", f"{cache_base}/.cache")
 
         # Ensure directories exist
         for env_var in [
-            "TORCH_HOME",
             "MPLCONFIGDIR",
             "FONTCONFIG_CACHE",
             "XDG_CACHE_HOME",
@@ -89,10 +94,9 @@ def setup_hpc_environment() -> None:
 
     # WandB configuration (offline by default for HPC)
     os.environ.setdefault("WANDB_MODE", "offline")
-    if cache_base:
-        os.environ.setdefault("WANDB_DIR", f"{cache_base}/.wandb")
-        os.environ.setdefault("WANDB_CACHE_DIR", f"{cache_base}/.wandb_cache")
-        os.environ.setdefault("WANDB_CONFIG_DIR", f"{cache_base}/.wandb_config")
+    os.environ.setdefault("WANDB_DIR", f"{cache_base}/.wandb")
+    os.environ.setdefault("WANDB_CACHE_DIR", f"{cache_base}/.wandb_cache")
+    os.environ.setdefault("WANDB_CONFIG_DIR", f"{cache_base}/.wandb_config")
 
     # Suppress non-critical warnings
     os.environ.setdefault(

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/models/resnet.py

@@ -49,6 +49,36 @@ def _get_conv_layers(
         raise ValueError(f"Unsupported dimensionality: {dim}D. Supported: 1D, 2D, 3D.")
 
 
+def _get_num_groups(num_channels: int, preferred_groups: int = 32) -> int:
+    """
+    Get valid num_groups for GroupNorm that divides num_channels evenly.
+
+    Args:
+        num_channels: Number of channels to normalize
+        preferred_groups: Preferred number of groups (default: 32)
+
+    Returns:
+        Valid num_groups that divides num_channels
+
+    Raises:
+        ValueError: If no valid divisor found (shouldn't happen with power-of-2 channels)
+    """
+    # Try preferred groups first, then decrease
+    for groups in [preferred_groups, 16, 8, 4, 2, 1]:
+        if groups <= num_channels and num_channels % groups == 0:
+            return groups
+
+    # Fallback: find any valid divisor
+    for groups in range(min(32, num_channels), 0, -1):
+        if num_channels % groups == 0:
+            return groups
+
+    raise ValueError(
+        f"Cannot find valid num_groups for {num_channels} channels. "
+        f"Consider using base_width that is a power of 2 (e.g., 32, 64, 128)."
+    )
+
+
 class BasicBlock(nn.Module):
     """
     Basic residual block for ResNet-18/34.
@@ -77,12 +107,12 @@ class BasicBlock(nn.Module):
             padding=1,
             bias=False,
         )
-        self.gn1 = nn.GroupNorm(min(32, out_channels), out_channels)
+        self.gn1 = nn.GroupNorm(_get_num_groups(out_channels), out_channels)
         self.relu = nn.ReLU(inplace=True)
         self.conv2 = Conv(
             out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False
         )
-        self.gn2 = nn.GroupNorm(min(32, out_channels), out_channels)
+        self.gn2 = nn.GroupNorm(_get_num_groups(out_channels), out_channels)
         self.downsample = downsample
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -125,7 +155,7 @@ class Bottleneck(nn.Module):
 
         # 1x1 reduce
         self.conv1 = Conv(in_channels, out_channels, kernel_size=1, bias=False)
-        self.gn1 = nn.GroupNorm(min(32, out_channels), out_channels)
+        self.gn1 = nn.GroupNorm(_get_num_groups(out_channels), out_channels)
 
         # 3x3 conv
         self.conv2 = Conv(
@@ -136,15 +166,14 @@ class Bottleneck(nn.Module):
             padding=1,
             bias=False,
         )
-        self.gn2 = nn.GroupNorm(min(32, out_channels), out_channels)
+        self.gn2 = nn.GroupNorm(_get_num_groups(out_channels), out_channels)
 
         # 1x1 expand
         self.conv3 = Conv(
             out_channels, out_channels * self.expansion, kernel_size=1, bias=False
         )
-        self.gn3 = nn.GroupNorm(
-            min(32, out_channels * self.expansion), out_channels * self.expansion
-        )
+        expanded_channels = out_channels * self.expansion
+        self.gn3 = nn.GroupNorm(_get_num_groups(expanded_channels), expanded_channels)
 
         self.relu = nn.ReLU(inplace=True)
         self.downsample = downsample
@@ -200,7 +229,7 @@ class ResNetBase(BaseModel):
 
         # Stem: 7x7 conv (or equivalent for 1D/3D)
         self.conv1 = Conv(1, base_width, kernel_size=7, stride=2, padding=3, bias=False)
-        self.gn1 = nn.GroupNorm(min(32, base_width), base_width)
+        self.gn1 = nn.GroupNorm(_get_num_groups(base_width), base_width)
         self.relu = nn.ReLU(inplace=True)
         self.maxpool = MaxPool(kernel_size=3, stride=2, padding=1)
 
@@ -246,7 +275,7 @@ class ResNetBase(BaseModel):
                     bias=False,
                 ),
                 nn.GroupNorm(
-                    min(32, out_channels * block.expansion),
+                    _get_num_groups(out_channels * block.expansion),
                     out_channels * block.expansion,
                 ),
             )

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/train.py

@@ -69,6 +69,39 @@ _setup_cache_dir("XDG_DATA_HOME", "local/share")
 _setup_cache_dir("XDG_STATE_HOME", "local/state")
 _setup_cache_dir("XDG_CACHE_HOME", "cache")
 
+
+def _setup_per_rank_compile_cache() -> None:
+    """Set per-GPU Triton/Inductor cache to prevent multi-process race warnings.
+
+    When using torch.compile with multiple GPUs, all processes try to write to
+    the same cache directory, causing 'Directory is not empty - skipping!' warnings.
+    This gives each GPU rank its own isolated cache subdirectory.
+    """
+    # Get local rank from environment (set by accelerate/torchrun)
+    local_rank = os.environ.get("LOCAL_RANK", "0")
+
+    # Get cache base from environment or use CWD
+    cache_base = os.environ.get(
+        "TRITON_CACHE_DIR", os.path.join(os.getcwd(), ".triton_cache")
+    )
+
+    # Set per-rank cache directories
+    os.environ["TRITON_CACHE_DIR"] = os.path.join(cache_base, f"rank_{local_rank}")
+    os.environ["TORCHINDUCTOR_CACHE_DIR"] = os.path.join(
+        os.environ.get(
+            "TORCHINDUCTOR_CACHE_DIR", os.path.join(os.getcwd(), ".inductor_cache")
+        ),
+        f"rank_{local_rank}",
+    )
+
+    # Create directories
+    os.makedirs(os.environ["TRITON_CACHE_DIR"], exist_ok=True)
+    os.makedirs(os.environ["TORCHINDUCTOR_CACHE_DIR"], exist_ok=True)
+
+
+# Setup per-rank compile caches (before torch imports)
+_setup_per_rank_compile_cache()
+
 # =============================================================================
 # Standard imports (after environment setup)
 # =============================================================================
@@ -908,7 +941,6 @@ def main():
         logger.info("=" * len(header))
 
     try:
-        time.time()
         total_training_time = 0.0
 
         for epoch in range(start_epoch, args.epochs):
@@ -1002,49 +1034,29 @@ def main():
                     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)
-
-            # Gather predictions and targets to rank 0 only (memory-efficient)
-            # Avoids duplicating full validation set on every GPU
-            if torch.distributed.is_initialized():
-                # DDP mode: gather only to rank 0
-                # NCCL backend requires CUDA tensors for collective ops
-                gpu_preds = cpu_preds.to(accelerator.device)
-                gpu_targets = cpu_targets.to(accelerator.device)
-
-                if accelerator.is_main_process:
-                    # Rank 0: allocate gather buffers on GPU
-                    all_preds_list = [
-                        torch.zeros_like(gpu_preds)
-                        for _ in range(accelerator.num_processes)
-                    ]
-                    all_targets_list = [
-                        torch.zeros_like(gpu_targets)
-                        for _ in range(accelerator.num_processes)
-                    ]
-                    torch.distributed.gather(
-                        gpu_preds, gather_list=all_preds_list, dst=0
-                    )
-                    torch.distributed.gather(
-                        gpu_targets, gather_list=all_targets_list, dst=0
-                    )
-                    # Move back to CPU for metric computation
-                    gathered = [
-                        (
-                            torch.cat(all_preds_list).cpu(),
-                            torch.cat(all_targets_list).cpu(),
-                        )
-                    ]
-                else:
-                    # Other ranks: send to rank 0, don't allocate gather buffers
-                    torch.distributed.gather(gpu_preds, gather_list=None, dst=0)
-                    torch.distributed.gather(gpu_targets, gather_list=None, dst=0)
-                    gathered = [(cpu_preds, cpu_targets)]  # Placeholder, not used
+            # Concatenate locally (keep on GPU for gather_for_metrics compatibility)
+            local_preds_cat = torch.cat(local_preds)
+            local_targets_cat = torch.cat(local_targets)
+
+            # Gather predictions and targets using Accelerate's CPU-efficient utility
+            # gather_for_metrics handles:
+            # - DDP padding removal (no need to trim manually)
+            # - Efficient cross-rank gathering without GPU memory spike
+            # - Returns concatenated tensors on CPU for metric computation
+            if accelerator.num_processes > 1:
+                # Move to GPU for gather (required by NCCL), then back to CPU
+                # gather_for_metrics is more memory-efficient than manual gather
+                # as it processes in chunks internally
+                gathered_preds = accelerator.gather_for_metrics(
+                    local_preds_cat.to(accelerator.device)
+                ).cpu()
+                gathered_targets = accelerator.gather_for_metrics(
+                    local_targets_cat.to(accelerator.device)
+                ).cpu()
             else:
                 # Single-GPU mode: no gathering needed
-                gathered = [(cpu_preds, cpu_targets)]
+                gathered_preds = local_preds_cat
+                gathered_targets = local_targets_cat
 
             # Synchronize validation metrics (scalars only - efficient)
             val_loss_scalar = val_loss_sum.item()
@@ -1069,20 +1081,10 @@ def main():
 
             # ==================== LOGGING & CHECKPOINTING ====================
             if accelerator.is_main_process:
-                # 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)
-                if len(y_pred) > real_len:
-                    y_pred = y_pred[:real_len]
-                    y_true = y_true[:real_len]
+                # gather_for_metrics already handles DDP padding removal
+                y_pred = gathered_preds.float().numpy()
+                y_true = gathered_targets.float().numpy()
 
                 # Guard against tiny validation sets (R² undefined for <2 samples)
                 if len(y_true) >= 2:

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/utils/config.py

@@ -183,9 +183,11 @@ def save_config(
             config[key] = value
 
     # Add metadata
+    from wavedl import __version__
+
     config["_metadata"] = {
         "saved_at": datetime.now().isoformat(),
-        "wavedl_version": "1.0.0",
+        "wavedl_version": __version__,
     }
 
     output_path = Path(output_path)

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/utils/cross_validation.py

@@ -337,6 +337,17 @@ def run_cross_validation(
         torch.cuda.manual_seed_all(seed)
 
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    # Auto-detect optimal DataLoader workers if not specified (matches train.py behavior)
+    if workers < 0:
+        cpu_count = os.cpu_count() or 4
+        num_gpus = torch.cuda.device_count() if torch.cuda.is_available() else 1
+        # Heuristic: 4-16 workers per GPU, bounded by available CPU cores
+        workers = min(16, max(2, (cpu_count - 2) // max(1, num_gpus)))
+        logger.info(
+            f"⚙️  Auto-detected workers: {workers} (CPUs: {cpu_count}, GPUs: {num_gpus})"
+        )
+
     logger.info(f"🚀 K-Fold Cross-Validation ({folds} folds)")
     logger.info(f"   Model: {model_name} | Device: {device}")
     logger.info(

{wavedl-1.5.1 → wavedl-1.5.3}/src/wavedl/utils/data.py

@@ -202,18 +202,31 @@ class NPZSource(DataSource):
     """Load data from NumPy .npz archives."""
 
     @staticmethod
-    def _safe_load(path: str, mmap_mode: str | None = None):
-        """Load NPZ with pickle only if needed (sparse matrix support)."""
+    def _safe_load(path: str, keys_to_probe: list[str], mmap_mode: str | None = None):
+        """Load NPZ with pickle only if needed (sparse matrix support).
+
+        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.
+        """
+        data = np.load(path, allow_pickle=False, mmap_mode=mmap_mode)
         try:
-            return np.load(path, allow_pickle=False, mmap_mode=mmap_mode)
-        except ValueError:
-            # Fallback for sparse matrices stored as object arrays
-            return np.load(path, allow_pickle=True, mmap_mode=mmap_mode)
+            # Probe keys to trigger error if object arrays exist
+            for key in keys_to_probe:
+                if key in data:
+                    _ = data[key]  # This raises ValueError for object arrays
+            return data
+        except ValueError as e:
+            if "allow_pickle=False" in str(e):
+                # Fallback for sparse matrices stored as object arrays
+                data.close() if hasattr(data, "close") else None
+                return np.load(path, allow_pickle=True, mmap_mode=mmap_mode)
+            raise
 
     def load(self, path: str) -> tuple[np.ndarray, np.ndarray]:
         """Load NPZ file (pickle enabled only for sparse matrices)."""
-        data = self._safe_load(path)
-        keys = list(data.keys())
+        # First pass to find keys without loading data
+        with np.load(path, allow_pickle=False) as probe:
+            keys = list(probe.keys())
 
         input_key = self._find_key(keys, INPUT_KEYS)
         output_key = self._find_key(keys, OUTPUT_KEYS)
@@ -225,6 +238,7 @@ class NPZSource(DataSource):
                 f"Found: {keys}"
             )
 
+        data = self._safe_load(path, [input_key, output_key])
         inp = data[input_key]
         outp = data[output_key]
 
@@ -243,8 +257,9 @@ class NPZSource(DataSource):
 
         Note: Returns memory-mapped arrays - do NOT modify them.
         """
-        data = self._safe_load(path, mmap_mode="r")
-        keys = list(data.keys())
+        # First pass to find keys without loading data
+        with np.load(path, allow_pickle=False) as probe:
+            keys = list(probe.keys())
 
         input_key = self._find_key(keys, INPUT_KEYS)
         output_key = self._find_key(keys, OUTPUT_KEYS)
@@ -256,6 +271,7 @@ class NPZSource(DataSource):
                 f"Found: {keys}"
             )
 
+        data = self._safe_load(path, [input_key, output_key], mmap_mode="r")
         inp = data[input_key]
         outp = data[output_key]
 
@@ -263,8 +279,9 @@ class NPZSource(DataSource):
 
     def load_outputs_only(self, path: str) -> np.ndarray:
         """Load only targets from NPZ (avoids loading large input arrays)."""
-        data = self._safe_load(path)
-        keys = list(data.keys())
+        # First pass to find keys without loading data
+        with np.load(path, allow_pickle=False) as probe:
+            keys = list(probe.keys())
 
         output_key = self._find_key(keys, OUTPUT_KEYS)
         if output_key is None:
@@ -273,6 +290,7 @@ class NPZSource(DataSource):
                 f"Supported keys: {OUTPUT_KEYS}. Found: {keys}"
             )
 
+        data = self._safe_load(path, [output_key])
         return data[output_key]
 
 
@@ -745,25 +763,77 @@ def load_test_data(
             k for k in OUTPUT_KEYS if k != "output_test"
         ]
 
-    # Load data using appropriate source
+    # Load data using appropriate source with test-key priority
+    # We detect keys first to ensure input_test/output_test are used when present
     try:
-        inp, outp = source.load(path)
+        if format == "npz":
+            with np.load(path, allow_pickle=False) as probe:
+                keys = list(probe.keys())
+            inp_key = DataSource._find_key(keys, custom_input_keys)
+            out_key = DataSource._find_key(keys, custom_output_keys)
+            if inp_key is None:
+                raise KeyError(
+                    f"Input key not found. Tried: {custom_input_keys}. Found: {keys}"
+                )
+            data = NPZSource._safe_load(
+                path, [inp_key] + ([out_key] if out_key else [])
+            )
+            inp = data[inp_key]
+            if inp.dtype == object:
+                inp = np.array(
+                    [x.toarray() if hasattr(x, "toarray") else x for x in inp]
+                )
+            outp = data[out_key] if out_key else None
+        elif format == "hdf5":
+            with h5py.File(path, "r") as f:
+                keys = list(f.keys())
+                inp_key = DataSource._find_key(keys, custom_input_keys)
+                out_key = DataSource._find_key(keys, custom_output_keys)
+                if inp_key is None:
+                    raise KeyError(
+                        f"Input key not found. Tried: {custom_input_keys}. Found: {keys}"
+                    )
+                inp = f[inp_key][:]
+                outp = f[out_key][:] if out_key else None
+        elif format == "mat":
+            mat_source = MATSource()
+            with h5py.File(path, "r") as f:
+                keys = list(f.keys())
+                inp_key = DataSource._find_key(keys, custom_input_keys)
+                out_key = DataSource._find_key(keys, custom_output_keys)
+                if inp_key is None:
+                    raise KeyError(
+                        f"Input key not found. Tried: {custom_input_keys}. Found: {keys}"
+                    )
+                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
+                else:
+                    outp = None
+        else:
+            # Fallback to default source.load() for unknown formats
+            inp, outp = source.load(path)
     except KeyError:
         # Try with just inputs if outputs not found (inference-only mode)
         if format == "npz":
-            data = NPZSource._safe_load(path)
-            keys = list(data.keys())
+            # First pass to find keys
+            with np.load(path, allow_pickle=False) as probe:
+                keys = list(probe.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}"
                 )
+            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)
             inp = data[inp_key]
             if inp.dtype == object:
                 inp = np.array(
                     [x.toarray() if hasattr(x, "toarray") else x for x in inp]
                 )
-            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

{wavedl-1.5.1 → wavedl-1.5.3/src/wavedl.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: wavedl
-Version: 1.5.1
+Version: 1.5.3
 Summary: A Scalable Deep Learning Framework for Wave-Based Inverse Problems
 Author: Ductho Le
 License: MIT
@@ -301,8 +301,8 @@ python -m wavedl.test --checkpoint <checkpoint_folder> --data_path <test_data> \
 
 **Requirements** (your model must):
 1. Inherit from `BaseModel`
-2. Accept `in_channels`, `num_outputs`, `input_shape` in `__init__`
-3. Return a tensor of shape `(batch, num_outputs)` from `forward()`
+2. Accept `in_shape`, `out_size` in `__init__`
+3. Return a tensor of shape `(batch, out_size)` from `forward()`
 
 ---
 
@@ -315,23 +315,22 @@ from wavedl.models import BaseModel, register_model
 
 @register_model("my_model")  # This name is used with --model flag
 class MyModel(BaseModel):
-    def __init__(self, in_channels, num_outputs, input_shape):
-        # in_channels: number of input channels (auto-detected from data)
-        # num_outputs: number of parameters to predict (auto-detected from data)
-        # input_shape: spatial dimensions, e.g., (128,) or (64, 64) or (32, 32, 32)
-        super().__init__(in_channels, num_outputs, input_shape)
-
-        # Define your layers (this is just an example)
-        self.conv1 = nn.Conv2d(in_channels, 64, 3, padding=1)
+    def __init__(self, in_shape, out_size, **kwargs):
+        # in_shape: spatial dimensions, e.g., (128,) or (64, 64) or (32, 32, 32)
+        # out_size: number of parameters to predict (auto-detected from data)
+        super().__init__(in_shape, out_size)
+
+        # Define your layers (this is just an example for 2D)
+        self.conv1 = nn.Conv2d(1, 64, 3, padding=1)  # Input always has 1 channel
         self.conv2 = nn.Conv2d(64, 128, 3, padding=1)
-        self.fc = nn.Linear(128, num_outputs)
+        self.fc = nn.Linear(128, out_size)
 
     def forward(self, x):
-        # Input x has shape: (batch, in_channels, *input_shape)
+        # Input x has shape: (batch, 1, *in_shape)
         x = F.relu(self.conv1(x))
         x = F.relu(self.conv2(x))
         x = x.mean(dim=[-2, -1])  # Global average pooling
-        return self.fc(x)  # Output shape: (batch, num_outputs)
+        return self.fc(x)  # Output shape: (batch, out_size)
 ```
 
 **Step 2: Train**
@@ -573,14 +572,19 @@ WaveDL automatically enables performance optimizations for modern GPUs:
 </details>
 
 <details>
-<summary><b>Environment Variables (wavedl-hpc)</b></summary>
+<summary><b>HPC CLI Arguments (wavedl-hpc)</b></summary>
+
+| Argument | Default | Description |
+|----------|---------|-------------|
+| `--num_gpus` | **Auto-detected** | Number of GPUs to use. By default, automatically detected via `nvidia-smi`. Set explicitly to override |
+| `--num_machines` | `1` | Number of machines in distributed setup |
+| `--mixed_precision` | `bf16` | Precision mode: `bf16`, `fp16`, or `no` |
+| `--dynamo_backend` | `no` | PyTorch Dynamo backend |
+
+**Environment Variables (for logging):**
 
 | Variable | Default | Description |
 |----------|---------|-------------|
-| `NUM_GPUS` | **Auto-detected** | Number of GPUs to use. By default, automatically detected via `nvidia-smi`. Set explicitly to override (e.g., `NUM_GPUS=2`) |
-| `NUM_MACHINES` | `1` | Number of machines in distributed setup |
-| `MIXED_PRECISION` | `bf16` | Precision mode: `bf16`, `fp16`, or `no` |
-| `DYNAMO_BACKEND` | `no` | PyTorch Dynamo backend |
 | `WANDB_MODE` | `offline` | WandB mode: `offline` or `online` |
 
 </details>