wavedl 1.5.5__tar.gz → 1.5.7__tar.gz

{wavedl-1.5.5/src/wavedl.egg-info → wavedl-1.5.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: wavedl
-Version: 1.5.5
+Version: 1.5.7
 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 (731 tests)
+├── unit_tests/                   # Pytest test suite (903 tests)
 │
 ├── pyproject.toml                # Package config, dependencies
 ├── CHANGELOG.md                  # Version history
@@ -470,6 +470,7 @@ WaveDL/
 ⭐ = **Pretrained on ImageNet** (recommended for smaller datasets). Weights are downloaded automatically on first use.
 - **Cache location**: `~/.cache/torch/hub/checkpoints/` (or `./.torch_cache/` on HPC if home is not writable)
 - **Size**: ~20–350 MB per model depending on architecture
+- **Train from scratch**: Use `--no_pretrained` to disable pretrained weights
 
 **💡 HPC Users**: If compute nodes block internet, pre-download weights on the login node:
 
@@ -1030,37 +1031,46 @@ print(f"✓ Output: {data['output_train'].shape} {data['output_train'].dtype}")
 
 ## 📦 Examples [![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)
 
-The `examples/` folder contains a **complete, ready-to-run example** for **material characterization of isotropic plates**. The pre-trained CNN predicts three physical parameters from Lamb wave dispersion curves:
+The `examples/` folder contains a **complete, ready-to-run example** for **material characterization of isotropic plates**. The pre-trained MobileNetV3 predicts three physical parameters from Lamb wave dispersion curves:
 
 | Parameter | Unit | Description |
 |-----------|------|-------------|
-| *h* | mm | Plate thickness |
-| √(*E*/ρ) | km/s | Square root of Young's modulus over density |
-| *ν* | — | Poisson's ratio |
+| $h$ | mm | Plate thickness |
+| $\sqrt{E/\rho}$ | km/s | Square root of Young's modulus over density |
+| $\nu$ | — | Poisson's ratio |
 
 > [!NOTE]
-> This example is based on our paper at **SPIE Smart Structures + NDE 2026**: [*"Deep learning-based ultrasonic assessment of plate thickness and elasticity"*](https://spie.org/spie-smart-structures-and-materials-nondestructive-evaluation/presentation/Deep-learningbased-ultrasonic-assessment-of-plate-thickness-and-elasticity/13951-4) (Paper 13951-4, to appear).
+> This example is based on our paper at **SPIE Smart Structures + NDE 2026**: [*"A lightweight deep learning model for ultrasonic assessment of plate thickness and elasticity
+"*](https://spie.org/spie-smart-structures-and-materials-nondestructive-evaluation/presentation/A-lightweight-deep-learning-model-for-ultrasonic-assessment-of-plate/13951-4) (Paper 13951-4, to appear).
+
+**Sample Dispersion Data:**
+
+<p align="center">
+  <img src="examples/elasticity_prediction/dispersion_samples.png" alt="Dispersion curve samples" width="700"><br>
+  <em>Test samples showing the wavenumber-frequency relationship for different plate properties</em>
+</p>
 
 **Try it yourself:**
 
 ```bash
 # Run inference on the example data
-python -m wavedl.test --checkpoint ./examples/elastic_cnn_example/best_checkpoint \
-  --data_path ./examples/elastic_cnn_example/Test_data_500.mat \
-  --plot --save_predictions --output_dir ./examples/elastic_cnn_example/test_results
+python -m wavedl.test --checkpoint ./examples/elasticity_prediction/best_checkpoint \
+  --data_path ./examples/elasticity_prediction/Test_data_100.mat \
+  --plot --save_predictions --output_dir ./examples/elasticity_prediction/test_results
 
 # Export to ONNX (already included as model.onnx)
-python -m wavedl.test --checkpoint ./examples/elastic_cnn_example/best_checkpoint \
-  --data_path ./examples/elastic_cnn_example/Test_data_500.mat \
-  --export onnx --export_path ./examples/elastic_cnn_example/model.onnx
+python -m wavedl.test --checkpoint ./examples/elasticity_prediction/best_checkpoint \
+  --data_path ./examples/elasticity_prediction/Test_data_100.mat \
+  --export onnx --export_path ./examples/elasticity_prediction/model.onnx
 ```
 
 **What's Included:**
 
 | File | Description |
 |------|-------------|
-| `best_checkpoint/` | Pre-trained CNN checkpoint |
-| `Test_data_500.mat` | 500 sample test set (500×500 dispersion curves → *h*, √(*E*/ρ), *ν*) |
+| `best_checkpoint/` | Pre-trained MobileNetV3 checkpoint |
+| `Test_data_100.mat` | 100 sample test set (500×500 dispersion curves → $h$, $\sqrt{E/\rho}$, $\nu$) |
+| `dispersion_samples.png` | Visualization of sample dispersion curves with material parameters |
 | `model.onnx` | ONNX export with embedded de-normalization |
 | `training_history.csv` | Epoch-by-epoch training metrics (loss, R², LR, etc.) |
 | `training_curves.png` | Training/validation loss and learning rate plot |
@@ -1070,59 +1080,59 @@ python -m wavedl.test --checkpoint ./examples/elastic_cnn_example/best_checkpoin
 **Training Progress:**
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/training_curves.png" alt="Training curves" width="600"><br>
-  <em>Training and validation loss over 227 epochs with <code>onecycle</code> learning rate schedule</em>
+  <img src="examples/elasticity_prediction/training_curves.png" alt="Training curves" width="600"><br>
+  <em>Training and validation loss with <code>plateau</code> learning rate schedule</em>
 </p>
 
 **Inference Results:**
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/scatter_all.png" alt="Scatter plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/scatter_all.png" alt="Scatter plot" width="700"><br>
   <em>Figure 1: Predictions vs ground truth for all three elastic parameters</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_histogram.png" alt="Error histogram" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/error_histogram.png" alt="Error histogram" width="700"><br>
   <em>Figure 2: Distribution of prediction errors showing near-zero mean bias</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/residuals.png" alt="Residual plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/residuals.png" alt="Residual plot" width="700"><br>
   <em>Figure 3: Residuals vs predicted values (no heteroscedasticity detected)</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/bland_altman.png" alt="Bland-Altman plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/bland_altman.png" alt="Bland-Altman plot" width="700"><br>
   <em>Figure 4: Bland-Altman analysis with ±1.96 SD limits of agreement</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/qq_plot.png" alt="Q-Q plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/qq_plot.png" alt="Q-Q plot" width="700"><br>
   <em>Figure 5: Q-Q plots confirming normally distributed prediction errors</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_correlation.png" alt="Error correlation" width="300"><br>
+  <img src="examples/elasticity_prediction/test_results/error_correlation.png" alt="Error correlation" width="300"><br>
   <em>Figure 6: Error correlation matrix between parameters</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/relative_error.png" alt="Relative error" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/relative_error.png" alt="Relative error" width="700"><br>
   <em>Figure 7: Relative error (%) vs true value for each parameter</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_cdf.png" alt="Error CDF" width="500"><br>
+  <img src="examples/elasticity_prediction/test_results/error_cdf.png" alt="Error CDF" width="500"><br>
   <em>Figure 8: Cumulative error distribution — 95% of predictions within indicated bounds</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/prediction_vs_index.png" alt="Prediction vs index" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/prediction_vs_index.png" alt="Prediction vs index" width="700"><br>
   <em>Figure 9: True vs predicted values by sample index</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_boxplot.png" alt="Error box plot" width="400"><br>
+  <img src="examples/elasticity_prediction/test_results/error_boxplot.png" alt="Error box plot" width="400"><br>
   <em>Figure 10: Error distribution summary (median, quartiles, outliers)</em>
 </p>
 

{wavedl-1.5.5 → wavedl-1.5.7}/README.md

@@ -342,7 +342,7 @@ WaveDL/
 ├── configs/                      # YAML config templates
 ├── examples/                     # Ready-to-run examples
 ├── notebooks/                    # Jupyter notebooks
-├── unit_tests/                   # Pytest test suite (731 tests)
+├── unit_tests/                   # Pytest test suite (903 tests)
 │
 ├── pyproject.toml                # Package config, dependencies
 ├── CHANGELOG.md                  # Version history
@@ -424,6 +424,7 @@ WaveDL/
 ⭐ = **Pretrained on ImageNet** (recommended for smaller datasets). Weights are downloaded automatically on first use.
 - **Cache location**: `~/.cache/torch/hub/checkpoints/` (or `./.torch_cache/` on HPC if home is not writable)
 - **Size**: ~20–350 MB per model depending on architecture
+- **Train from scratch**: Use `--no_pretrained` to disable pretrained weights
 
 **💡 HPC Users**: If compute nodes block internet, pre-download weights on the login node:
 
@@ -984,37 +985,46 @@ print(f"✓ Output: {data['output_train'].shape} {data['output_train'].dtype}")
 
 ## 📦 Examples [![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)
 
-The `examples/` folder contains a **complete, ready-to-run example** for **material characterization of isotropic plates**. The pre-trained CNN predicts three physical parameters from Lamb wave dispersion curves:
+The `examples/` folder contains a **complete, ready-to-run example** for **material characterization of isotropic plates**. The pre-trained MobileNetV3 predicts three physical parameters from Lamb wave dispersion curves:
 
 | Parameter | Unit | Description |
 |-----------|------|-------------|
-| *h* | mm | Plate thickness |
-| √(*E*/ρ) | km/s | Square root of Young's modulus over density |
-| *ν* | — | Poisson's ratio |
+| $h$ | mm | Plate thickness |
+| $\sqrt{E/\rho}$ | km/s | Square root of Young's modulus over density |
+| $\nu$ | — | Poisson's ratio |
 
 > [!NOTE]
-> This example is based on our paper at **SPIE Smart Structures + NDE 2026**: [*"Deep learning-based ultrasonic assessment of plate thickness and elasticity"*](https://spie.org/spie-smart-structures-and-materials-nondestructive-evaluation/presentation/Deep-learningbased-ultrasonic-assessment-of-plate-thickness-and-elasticity/13951-4) (Paper 13951-4, to appear).
+> This example is based on our paper at **SPIE Smart Structures + NDE 2026**: [*"A lightweight deep learning model for ultrasonic assessment of plate thickness and elasticity
+"*](https://spie.org/spie-smart-structures-and-materials-nondestructive-evaluation/presentation/A-lightweight-deep-learning-model-for-ultrasonic-assessment-of-plate/13951-4) (Paper 13951-4, to appear).
+
+**Sample Dispersion Data:**
+
+<p align="center">
+  <img src="examples/elasticity_prediction/dispersion_samples.png" alt="Dispersion curve samples" width="700"><br>
+  <em>Test samples showing the wavenumber-frequency relationship for different plate properties</em>
+</p>
 
 **Try it yourself:**
 
 ```bash
 # Run inference on the example data
-python -m wavedl.test --checkpoint ./examples/elastic_cnn_example/best_checkpoint \
-  --data_path ./examples/elastic_cnn_example/Test_data_500.mat \
-  --plot --save_predictions --output_dir ./examples/elastic_cnn_example/test_results
+python -m wavedl.test --checkpoint ./examples/elasticity_prediction/best_checkpoint \
+  --data_path ./examples/elasticity_prediction/Test_data_100.mat \
+  --plot --save_predictions --output_dir ./examples/elasticity_prediction/test_results
 
 # Export to ONNX (already included as model.onnx)
-python -m wavedl.test --checkpoint ./examples/elastic_cnn_example/best_checkpoint \
-  --data_path ./examples/elastic_cnn_example/Test_data_500.mat \
-  --export onnx --export_path ./examples/elastic_cnn_example/model.onnx
+python -m wavedl.test --checkpoint ./examples/elasticity_prediction/best_checkpoint \
+  --data_path ./examples/elasticity_prediction/Test_data_100.mat \
+  --export onnx --export_path ./examples/elasticity_prediction/model.onnx
 ```
 
 **What's Included:**
 
 | File | Description |
 |------|-------------|
-| `best_checkpoint/` | Pre-trained CNN checkpoint |
-| `Test_data_500.mat` | 500 sample test set (500×500 dispersion curves → *h*, √(*E*/ρ), *ν*) |
+| `best_checkpoint/` | Pre-trained MobileNetV3 checkpoint |
+| `Test_data_100.mat` | 100 sample test set (500×500 dispersion curves → $h$, $\sqrt{E/\rho}$, $\nu$) |
+| `dispersion_samples.png` | Visualization of sample dispersion curves with material parameters |
 | `model.onnx` | ONNX export with embedded de-normalization |
 | `training_history.csv` | Epoch-by-epoch training metrics (loss, R², LR, etc.) |
 | `training_curves.png` | Training/validation loss and learning rate plot |
@@ -1024,59 +1034,59 @@ python -m wavedl.test --checkpoint ./examples/elastic_cnn_example/best_checkpoin
 **Training Progress:**
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/training_curves.png" alt="Training curves" width="600"><br>
-  <em>Training and validation loss over 227 epochs with <code>onecycle</code> learning rate schedule</em>
+  <img src="examples/elasticity_prediction/training_curves.png" alt="Training curves" width="600"><br>
+  <em>Training and validation loss with <code>plateau</code> learning rate schedule</em>
 </p>
 
 **Inference Results:**
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/scatter_all.png" alt="Scatter plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/scatter_all.png" alt="Scatter plot" width="700"><br>
   <em>Figure 1: Predictions vs ground truth for all three elastic parameters</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_histogram.png" alt="Error histogram" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/error_histogram.png" alt="Error histogram" width="700"><br>
   <em>Figure 2: Distribution of prediction errors showing near-zero mean bias</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/residuals.png" alt="Residual plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/residuals.png" alt="Residual plot" width="700"><br>
   <em>Figure 3: Residuals vs predicted values (no heteroscedasticity detected)</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/bland_altman.png" alt="Bland-Altman plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/bland_altman.png" alt="Bland-Altman plot" width="700"><br>
   <em>Figure 4: Bland-Altman analysis with ±1.96 SD limits of agreement</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/qq_plot.png" alt="Q-Q plot" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/qq_plot.png" alt="Q-Q plot" width="700"><br>
   <em>Figure 5: Q-Q plots confirming normally distributed prediction errors</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_correlation.png" alt="Error correlation" width="300"><br>
+  <img src="examples/elasticity_prediction/test_results/error_correlation.png" alt="Error correlation" width="300"><br>
   <em>Figure 6: Error correlation matrix between parameters</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/relative_error.png" alt="Relative error" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/relative_error.png" alt="Relative error" width="700"><br>
   <em>Figure 7: Relative error (%) vs true value for each parameter</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_cdf.png" alt="Error CDF" width="500"><br>
+  <img src="examples/elasticity_prediction/test_results/error_cdf.png" alt="Error CDF" width="500"><br>
   <em>Figure 8: Cumulative error distribution — 95% of predictions within indicated bounds</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/prediction_vs_index.png" alt="Prediction vs index" width="700"><br>
+  <img src="examples/elasticity_prediction/test_results/prediction_vs_index.png" alt="Prediction vs index" width="700"><br>
   <em>Figure 9: True vs predicted values by sample index</em>
 </p>
 
 <p align="center">
-  <img src="examples/elastic_cnn_example/test_results/error_boxplot.png" alt="Error box plot" width="400"><br>
+  <img src="examples/elasticity_prediction/test_results/error_boxplot.png" alt="Error box plot" width="400"><br>
   <em>Figure 10: Error distribution summary (median, quartiles, outliers)</em>
 </p>
 

{wavedl-1.5.5 → wavedl-1.5.7}/src/wavedl/__init__.py

@@ -18,7 +18,7 @@ For inference:
     # or: python -m wavedl.test --checkpoint best_checkpoint --data_path test.npz
 """
 
-__version__ = "1.5.5"
+__version__ = "1.5.7"
 __author__ = "Ductho Le"
 __email__ = "ductho.le@outlook.com"
 

{wavedl-1.5.5 → wavedl-1.5.7}/src/wavedl/models/efficientnet.py

@@ -110,9 +110,30 @@ class EfficientNetBase(BaseModel):
             self._freeze_backbone()
 
     def _adapt_input_channels(self):
-        """Modify first conv to handle single-channel input by expanding to 3ch."""
-        # We'll handle this in forward by repeating channels
-        pass
+        """Modify first conv to accept single-channel input.
+
+        Instead of expanding 1→3 channels in forward (which triples memory),
+        we replace the first conv layer with a 1-channel version and initialize
+        weights as the mean of the pretrained RGB filters.
+        """
+        # EfficientNet stem conv is at: features[0][0]
+        old_conv = self.backbone.features[0][0]
+        new_conv = nn.Conv2d(
+            1,  # Single channel input
+            old_conv.out_channels,
+            kernel_size=old_conv.kernel_size,
+            stride=old_conv.stride,
+            padding=old_conv.padding,
+            dilation=old_conv.dilation,
+            groups=old_conv.groups,
+            padding_mode=old_conv.padding_mode,
+            bias=old_conv.bias is not None,
+        )
+        if self.pretrained:
+            # Initialize with mean of pretrained RGB weights
+            with torch.no_grad():
+                new_conv.weight.copy_(old_conv.weight.mean(dim=1, keepdim=True))
+        self.backbone.features[0][0] = new_conv
 
     def _freeze_backbone(self):
         """Freeze all backbone parameters except the classifier."""
@@ -130,10 +151,6 @@ class EfficientNetBase(BaseModel):
         Returns:
             Output tensor of shape (B, out_size)
         """
-        # Expand single channel to 3 channels for pretrained weights
-        if x.size(1) == 1:
-            x = x.expand(-1, 3, -1, -1)
-
         return self.backbone(x)
 
     @classmethod

{wavedl-1.5.5 → wavedl-1.5.7}/src/wavedl/models/efficientnetv2.py

@@ -129,10 +129,37 @@ class EfficientNetV2Base(BaseModel):
             nn.Linear(regression_hidden // 2, out_size),
         )
 
-        # Optionally freeze backbone for fine-tuning
+        # Adapt first conv for single-channel input (3× memory savings vs expand)
+        self._adapt_input_channels()
+
+        # Optionally freeze backbone for fine-tuning (after adaptation so new conv is frozen too)
         if freeze_backbone:
             self._freeze_backbone()
 
+    def _adapt_input_channels(self):
+        """Modify first conv to accept single-channel input.
+
+        Instead of expanding 1→3 channels in forward (which triples memory),
+        we replace the first conv layer with a 1-channel version and initialize
+        weights as the mean of the pretrained RGB filters.
+        """
+        old_conv = self.backbone.features[0][0]
+        new_conv = nn.Conv2d(
+            1,  # Single channel input
+            old_conv.out_channels,
+            kernel_size=old_conv.kernel_size,
+            stride=old_conv.stride,
+            padding=old_conv.padding,
+            dilation=old_conv.dilation,
+            groups=old_conv.groups,
+            padding_mode=old_conv.padding_mode,
+            bias=old_conv.bias is not None,
+        )
+        if self.pretrained:
+            with torch.no_grad():
+                new_conv.weight.copy_(old_conv.weight.mean(dim=1, keepdim=True))
+        self.backbone.features[0][0] = new_conv
+
     def _freeze_backbone(self):
         """Freeze all backbone parameters except the classifier."""
         for name, param in self.backbone.named_parameters():
@@ -144,15 +171,11 @@ class EfficientNetV2Base(BaseModel):
         Forward pass.
 
         Args:
-            x: Input tensor of shape (B, C, H, W) where C is 1 or 3
+            x: Input tensor of shape (B, 1, H, W)
 
         Returns:
             Output tensor of shape (B, out_size)
         """
-        # Expand single channel to 3 channels for pretrained weights compatibility
-        if x.size(1) == 1:
-            x = x.expand(-1, 3, -1, -1)
-
         return self.backbone(x)
 
     @classmethod

{wavedl-1.5.5 → wavedl-1.5.7}/src/wavedl/models/mobilenetv3.py

@@ -136,10 +136,37 @@ class MobileNetV3Base(BaseModel):
             nn.Linear(regression_hidden, out_size),
         )
 
-        # Optionally freeze backbone for fine-tuning
+        # Adapt first conv for single-channel input (3× memory savings vs expand)
+        self._adapt_input_channels()
+
+        # Optionally freeze backbone for fine-tuning (after adaptation so new conv is frozen too)
         if freeze_backbone:
             self._freeze_backbone()
 
+    def _adapt_input_channels(self):
+        """Modify first conv to accept single-channel input.
+
+        Instead of expanding 1→3 channels in forward (which triples memory),
+        we replace the first conv layer with a 1-channel version and initialize
+        weights as the mean of the pretrained RGB filters.
+        """
+        old_conv = self.backbone.features[0][0]
+        new_conv = nn.Conv2d(
+            1,  # Single channel input
+            old_conv.out_channels,
+            kernel_size=old_conv.kernel_size,
+            stride=old_conv.stride,
+            padding=old_conv.padding,
+            dilation=old_conv.dilation,
+            groups=old_conv.groups,
+            padding_mode=old_conv.padding_mode,
+            bias=old_conv.bias is not None,
+        )
+        if self.pretrained:
+            with torch.no_grad():
+                new_conv.weight.copy_(old_conv.weight.mean(dim=1, keepdim=True))
+        self.backbone.features[0][0] = new_conv
+
     def _freeze_backbone(self):
         """Freeze all backbone parameters except the classifier."""
         for name, param in self.backbone.named_parameters():
@@ -151,15 +178,11 @@ class MobileNetV3Base(BaseModel):
         Forward pass.
 
         Args:
-            x: Input tensor of shape (B, C, H, W) where C is 1 or 3
+            x: Input tensor of shape (B, 1, H, W)
 
         Returns:
             Output tensor of shape (B, out_size)
         """
-        # Expand single channel to 3 channels for pretrained weights compatibility
-        if x.size(1) == 1:
-            x = x.expand(-1, 3, -1, -1)
-
         return self.backbone(x)
 
     @classmethod
@@ -194,7 +217,7 @@ class MobileNetV3Small(MobileNetV3Base):
 
     Performance (approximate):
         - CPU inference: ~6ms (single core)
-        - Parameters: 2.5M
+        - Parameters: ~1.1M
         - MAdds: 56M
 
     Args:
@@ -241,7 +264,7 @@ class MobileNetV3Large(MobileNetV3Base):
 
     Performance (approximate):
         - CPU inference: ~20ms (single core)
-        - Parameters: 5.4M
+        - Parameters: ~3.2M
         - MAdds: 219M
 
     Args:

{wavedl-1.5.5 → wavedl-1.5.7}/src/wavedl/models/regnet.py

@@ -140,10 +140,37 @@ class RegNetBase(BaseModel):
             nn.Linear(regression_hidden, out_size),
         )
 
-        # Optionally freeze backbone for fine-tuning
+        # Adapt first conv for single-channel input (3× memory savings vs expand)
+        self._adapt_input_channels()
+
+        # Optionally freeze backbone for fine-tuning (after adaptation so new conv is frozen too)
         if freeze_backbone:
             self._freeze_backbone()
 
+    def _adapt_input_channels(self):
+        """Modify first conv to accept single-channel input.
+
+        Instead of expanding 1→3 channels in forward (which triples memory),
+        we replace the first conv layer with a 1-channel version and initialize
+        weights as the mean of the pretrained RGB filters.
+        """
+        old_conv = self.backbone.stem[0]
+        new_conv = nn.Conv2d(
+            1,  # Single channel input
+            old_conv.out_channels,
+            kernel_size=old_conv.kernel_size,
+            stride=old_conv.stride,
+            padding=old_conv.padding,
+            dilation=old_conv.dilation,
+            groups=old_conv.groups,
+            padding_mode=old_conv.padding_mode,
+            bias=old_conv.bias is not None,
+        )
+        if self.pretrained:
+            with torch.no_grad():
+                new_conv.weight.copy_(old_conv.weight.mean(dim=1, keepdim=True))
+        self.backbone.stem[0] = new_conv
+
     def _freeze_backbone(self):
         """Freeze all backbone parameters except the fc layer."""
         for name, param in self.backbone.named_parameters():
@@ -155,15 +182,11 @@ class RegNetBase(BaseModel):
         Forward pass.
 
         Args:
-            x: Input tensor of shape (B, C, H, W) where C is 1 or 3
+            x: Input tensor of shape (B, 1, H, W)
 
         Returns:
             Output tensor of shape (B, out_size)
         """
-        # Expand single channel to 3 channels for pretrained weights compatibility
-        if x.size(1) == 1:
-            x = x.expand(-1, 3, -1, -1)
-
         return self.backbone(x)
 
     @classmethod

{wavedl-1.5.5 → wavedl-1.5.7}/src/wavedl/models/swin.py

@@ -141,10 +141,46 @@ class SwinTransformerBase(BaseModel):
             nn.Linear(regression_hidden // 2, out_size),
         )
 
-        # Optionally freeze backbone for fine-tuning
+        # Adapt patch embedding conv for single-channel input (3× memory savings vs expand)
+        self._adapt_input_channels()
+
+        # Optionally freeze backbone for fine-tuning (after adaptation so new conv is frozen too)
         if freeze_backbone:
             self._freeze_backbone()
 
+    def _adapt_input_channels(self):
+        """Modify patch embedding conv to accept single-channel input.
+
+        Instead of expanding 1→3 channels in forward (which triples memory),
+        we replace the patch embedding conv with a 1-channel version and
+        initialize weights as the mean of the pretrained RGB filters.
+        """
+        # Swin's patch embedding is at features[0][0]
+        try:
+            old_conv = self.backbone.features[0][0]
+        except (IndexError, AttributeError, TypeError) as e:
+            raise RuntimeError(
+                f"Swin patch embed structure changed in this torchvision version. "
+                f"Cannot adapt input channels. Error: {e}"
+            ) from e
+        new_conv = nn.Conv2d(
+            1,  # Single channel input
+            old_conv.out_channels,
+            kernel_size=old_conv.kernel_size,
+            stride=old_conv.stride,
+            padding=old_conv.padding,
+            dilation=old_conv.dilation,
+            groups=old_conv.groups,
+            padding_mode=old_conv.padding_mode,
+            bias=old_conv.bias is not None,
+        )
+        if self.pretrained:
+            with torch.no_grad():
+                new_conv.weight.copy_(old_conv.weight.mean(dim=1, keepdim=True))
+                if old_conv.bias is not None:
+                    new_conv.bias.copy_(old_conv.bias)
+        self.backbone.features[0][0] = new_conv
+
     def _freeze_backbone(self):
         """Freeze all backbone parameters except the head."""
         for name, param in self.backbone.named_parameters():
@@ -156,15 +192,11 @@ class SwinTransformerBase(BaseModel):
         Forward pass.
 
         Args:
-            x: Input tensor of shape (B, C, H, W) where C is 1 or 3
+            x: Input tensor of shape (B, 1, H, W)
 
         Returns:
             Output tensor of shape (B, out_size)
         """
-        # Expand single channel to 3 channels for pretrained weights compatibility
-        if x.size(1) == 1:
-            x = x.expand(-1, 3, -1, -1)
-
         return self.backbone(x)
 
     @classmethod