feat: add neural additive model for binary classification

examples/binnam_tecator.py

@@ -0,0 +1,81 @@
+"""Neural Additive Model (NAM) example for binary classification."""
+
+import argparse
+
+import prototorch as pt
+import pytorch_lightning as pl
+import torch
+from matplotlib import pyplot as plt
+
+if __name__ == "__main__":
+    # Command-line arguments
+    parser = argparse.ArgumentParser()
+    parser = pl.Trainer.add_argparse_args(parser)
+    args = parser.parse_args()
+
+    # Dataset
+    train_ds = pt.datasets.Tecator("~/datasets")
+
+    # Dataloaders
+    train_loader = torch.utils.data.DataLoader(train_ds, batch_size=64)
+
+    # Hyperparameters
+    hparams = dict(lr=0.1)
+
+    # Define the feature extractor
+    class FE(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.modules_list = torch.nn.ModuleList([
+                torch.nn.Linear(1, 3),
+                torch.nn.Sigmoid(),
+                torch.nn.Linear(3, 1),
+                torch.nn.Sigmoid(),
+            ])
+
+        def forward(self, x):
+            for m in self.modules_list:
+                x = m(x)
+            return x
+
+    # Initialize the model
+    model = pt.models.BinaryNAM(
+        hparams,
+        extractors=torch.nn.ModuleList([FE() for _ in range(100)]),
+    )
+
+    # Compute intermediate input and output sizes
+    model.example_input_array = torch.zeros(4, 100)
+
+    # Callbacks
+    es = pl.callbacks.EarlyStopping(
+        monitor="train_loss",
+        min_delta=0.001,
+        patience=20,
+        mode="min",
+        verbose=True,
+        check_on_train_epoch_end=True,
+    )
+
+    # Setup trainer
+    trainer = pl.Trainer.from_argparse_args(
+        args,
+        callbacks=[
+            es,
+        ],
+        terminate_on_nan=True,
+        weights_summary=None,
+        accelerator="ddp",
+    )
+
+    # Training loop
+    trainer.fit(model, train_loader)
+
+    # Visualize extractor shape functions
+    fig, axes = plt.subplots(10, 10)
+    for i, ax in enumerate(axes.flat):
+        x = torch.linspace(-2, 2, 100)  # TODO use min/max from data
+        y = model.extractors[i](x.view(100, 1)).squeeze().detach()
+        ax.plot(x, y)
+        ax.set(title=f"Feature {i + 1}", xticklabels=[], yticklabels=[])
+    plt.show()

prototorch/models/__init__.py

@@ -19,6 +19,7 @@ from .glvq import (
 )
 from .knn import KNN
 from .lvq import LVQ1, LVQ21, MedianLVQ
+from .nam import BinaryNAM
 from .probabilistic import CELVQ, PLVQ, RSLVQ, SLVQ
 from .unsupervised import GrowingNeuralGas, HeskesSOM, KohonenSOM, NeuralGas
 from .vis import *

prototorch/models/nam.py

@@ -0,0 +1,44 @@
+"""ProtoTorch Neural Additive Model."""
+
+import torch
+import torchmetrics
+
+from .abstract import ProtoTorchBolt
+
+
+class BinaryNAM(ProtoTorchBolt):
+    """Neural Additive Model for binary classification.
+
+    Paper: https://arxiv.org/abs/2004.13912
+    Official implementation: https://github.com/google-research/google-research/tree/master/neural_additive_models
+
+    """
+    def __init__(self, hparams: dict, extractors: torch.nn.ModuleList,
+                 **kwargs):
+        super().__init__(hparams, **kwargs)
+        self.extractors = extractors
+
+    def extract(self, x):
+        """Apply the local extractors batch-wise on features."""
+        out = torch.zeros_like(x)
+        for j in range(x.shape[1]):
+            out[:, j] = self.extractors[j](x[:, j].unsqueeze(1)).squeeze()
+        return out
+
+    def forward(self, x):
+        x = self.extract(x).sum(1)
+        return torch.nn.functional.sigmoid(x)
+
+    def training_step(self, batch, batch_idx, optimizer_idx=None):
+        x, y = batch
+        preds = self(x)
+        train_loss = torch.nn.functional.binary_cross_entropy(preds, y.float())
+        self.log("train_loss", train_loss)
+        accuracy = torchmetrics.functional.accuracy(preds.int(), y.int())
+        self.log("train_acc",
+                 accuracy,
+                 on_step=False,
+                 on_epoch=True,
+                 prog_bar=True,
+                 logger=True)
+        return train_loss