Wrap FNO from NVIDIA Modulus.

Author: Samuel Burbulla

src/continuiti/operators/modulus/__init__.py

@@ -0,0 +1,17 @@
+"""
+`continuiti.operators.modulus`
+
+Operators from NVIDIA Modulus wrapped in continuiti.
+"""
+
+# Test if we can import NVIDIA Modulus
+try:
+    import modulus  # noqa: F40
+except ImportError:
+    raise ImportError("NVIDIA Modulus not found!")
+
+from .fno import FNO
+
+__all__ = [
+    "FNO",
+]

src/continuiti/operators/modulus/fno.py

@@ -0,0 +1,61 @@
+"""
+`continuiti.operators.modulus.fno`
+
+The Fourier Neural Operator from NVIDIA Modulus wrapped in continuiti.
+"""
+
+import torch
+from typing import Optional
+from continuiti.operators import Operator, OperatorShapes
+from modulus.models.fno import FNO as FNOModulus
+
+
+class FNO(Operator):
+    r"""FNO architecture from NVIDIA Modulus.
+
+    The `in_channels` and `out_channels` arguments are determined by the
+    `shapes` argument. The `dimension` is set to the dimension of the input
+    coordinates, assuming that the grid dimension is the same as the coordinate
+    dimension of `x`.
+
+    All other keyword arguments are passed to the Fourier Neural Operator, please refer
+    to the documentation of the `modulus.model.fno.FNO` class for more information.
+
+    Args:
+        shapes: Shapes of the input and output data.
+        device: Device.
+        **kwargs: Additional arguments for the Fourier layers.
+    """
+
+    def __init__(
+        self,
+        shapes: OperatorShapes,
+        device: Optional[torch.device] = None,
+        dimension: Optional[int] = None,
+        **kwargs,
+    ):
+        super().__init__(shapes, device)
+
+        if dimension is None:
+            # Per default, use coordinate dimension
+            dimension = shapes.x.dim
+
+        self.fno = FNOModulus(
+            in_channels=shapes.u.dim,
+            out_channels=shapes.v.dim,
+            dimension=dimension,
+            **kwargs,
+        )
+        self.fno.to(device)
+
+    def forward(
+        self, x: torch.Tensor, u: torch.Tensor, y: torch.Tensor
+    ) -> torch.Tensor:
+        r"""Forward pass of the Fourier Neural Operator.
+
+        Args:
+            x: Ignored.
+            u: Input function values of shape (batch_size, u_dim, num_sensors...).
+            y: Ignored.
+        """
+        return self.fno(u)

tests/operators/modulus/__init__.py

@@ -0,0 +1,52 @@
+import pytest
+from continuiti.benchmarks.sine import SineBenchmark
+from continuiti.trainer import Trainer
+from continuiti.operators.losses import MSELoss
+
+
+@pytest.mark.slow
+def test_modulus_fno():
+    try:
+        from continuiti.operators.modulus import FNO
+    except ImportError:
+        pytest.skip("NVIDIA Modulus not found!")
+
+    # Data set
+    benchmark = SineBenchmark(n_train=1)
+    dataset = benchmark.train_dataset
+
+    # Operator
+    #   Configured like the default continuiti `FourierNeuralOperator`
+    #   with depth=3 and width=3 as in `test_fno.py`.
+    operator = FNO(
+        dataset.shapes,
+        decoder_layers=1,
+        decoder_layer_size=1,
+        decoder_activation_fn="identity",
+        num_fno_layers=3,  # "depth" in FourierNeuralOperator
+        latent_channels=3,  # "width" in FourierNeuralOperator
+        num_fno_modes=dataset.shapes.u.size[0] // 2 + 1,
+        padding=0,
+        coord_features=False,
+    )
+
+    # Train
+    Trainer(operator, device="cpu").fit(dataset, tol=1e-12, epochs=10_000)
+
+    # Check solution
+    x, u, y, v = dataset.x, dataset.u, dataset.y, dataset.v
+    assert MSELoss()(operator, x, u, y, v) < 1e-12
+
+
+# SineBenchmark(n_train=1024, n_sensors=128, n_evaluations=128), epochs=100
+
+# NVIDIA Modulus FNO
+# Parameters: 3560  Device: cpu
+# Epoch 100/100  Step 32/32  [====================]  6ms/step  [0:19min<0:00min] - loss/train = 6.3876e-05
+
+# continuiti FNO
+# Parameters: 3556  Device: cpu
+# Epoch 100/100  Step 32/32  [====================]  3ms/step  [0:10min<0:00min] - loss/train = 1.4440e-04
+
+# -> continuiti FNO is 2x faster than NVIDIA Modulus FNO
+# -> NVIDIA Modulus FNO can not handle different number of sensors and evaluations