Add Averaging Neural Operator with tests and a tutorial (#230)

* add Averaging Neural Operator with tests
* add backward test
* minor changes
* doc addition

---------

Co-authored-by: Dario Coscia <dariocoscia@Dario-Coscia.local>

docs/source/_rst/_code.rst

@@ -56,6 +56,7 @@ Models
     MIONet <models/mionet.rst>
     FourierIntegralKernel <models/fourier_kernel.rst>
     FNO <models/fno.rst>
+    AveragingNeuralOperator <models/avno.rst>
 
 Layers
 -------------
@@ -67,11 +68,11 @@ Layers
     EnhancedLinear layer <layers/enhanced_linear.rst>
     Spectral convolution <layers/spectral.rst>
     Fourier layers <layers/fourier.rst>
+    Averaging layer <layers/avno_layer.rst>
     Continuous convolution <layers/convolution.rst>
     Proper Orthogonal Decomposition <layers/pod.rst>
     Periodic Boundary Condition embeddings <layers/embedding.rst>
 
-
 Equations and Operators
 -------------------------
 

docs/source/_rst/layers/avno_layer.rst

@@ -0,0 +1,8 @@
+Averaging layers
+====================
+.. currentmodule:: pina.model.layers.avno_layer
+
+.. autoclass:: AVNOBlock
+    :members:
+    :show-inheritance:
+    :noindex:

docs/source/_rst/models/avno.rst

@@ -0,0 +1,7 @@
+Averaging Neural Operator
+==============================
+.. currentmodule:: pina.model.avno
+
+.. autoclass:: AveragingNeuralOperator
+   :members:
+   :show-inheritance:

pina/model/__init__.py

@@ -7,6 +7,7 @@ __all__ = [
     "FNO",
     "FourierIntegralKernel",
     "KernelNeuralOperator",
+    "AveragingNeuralOperator",
 ]
 
 from .feed_forward import FeedForward, ResidualFeedForward
@@ -14,3 +15,4 @@ from .multi_feed_forward import MultiFeedForward
 from .deeponet import DeepONet, MIONet
 from .fno import FNO, FourierIntegralKernel
 from .base_no import KernelNeuralOperator
+from .avno import AveragingNeuralOperator

pina/model/avno.py

@@ -0,0 +1,104 @@
+"""Module Averaging Neural Operator."""
+
+from torch import nn, concatenate
+from . import FeedForward
+from .layers import AVNOBlock
+from .base_no import KernelNeuralOperator
+from pina.utils import check_consistency
+
+
+class AveragingNeuralOperator(KernelNeuralOperator):
+    """
+    Implementation of Averaging Neural Operator. 
+
+    Averaging Neural Operator is a general architecture for
+    learning Operators. Unlike traditional machine learning methods
+    AveragingNeuralOperator is designed to map entire functions
+    to other functions. It can be trained with Supervised learning strategies.
+    AveragingNeuralOperator does convolution by performing a field average.
+
+    .. seealso::
+
+        **Original reference**: Lanthaler S. Li, Z., Kovachki,
+        Stuart, A. (2020). *The Nonlocal Neural Operator:
+        Universal Approximation*.
+        DOI: `arXiv preprint arXiv:2304.13221.
+        <https://arxiv.org/abs/2304.13221>`_
+    """
+
+    def __init__(
+        self,
+        input_numb_fields,
+        output_numb_fields,
+        field_indices,
+        coordinates_indices,
+        dimension=3,
+        inner_size=100,
+        n_layers=4,
+        func=nn.GELU,
+    ):
+        """
+        :param int input_numb_fields: The number of input components 
+            of the model.
+        :param int output_numb_fields: The number of output components 
+            of the model.
+        :param int dimension: the dimension of the domain of the functions.
+        :param int inner_size: number of neurons in the hidden layer(s). 
+            Defaults to 100.
+        :param int n_layers: number of hidden layers. Default is 4.
+        :param func: the activation function to use. Default to nn.GELU.
+        :param list[str] field_indices: the label of the fields 
+            in the input tensor. 
+        :param list[str] coordinates_indices: the label of the 
+            coordinates in the input tensor. 
+        """
+
+        # check consistency
+        check_consistency(input_numb_fields, int)
+        check_consistency(output_numb_fields, int)
+        check_consistency(field_indices, str)
+        check_consistency(coordinates_indices, str)
+        check_consistency(dimension, int)
+        check_consistency(inner_size, int)
+        check_consistency(n_layers, int)
+        check_consistency(func, nn.Module, subclass=True)
+
+        # assign
+        self.input_numb_fields = input_numb_fields
+        self.output_numb_fields = output_numb_fields
+        self.dimension = dimension
+        self.coordinates_indices = coordinates_indices
+        self.field_indices = field_indices
+        integral_net = nn.Sequential(
+            *[AVNOBlock(inner_size, func) for _ in range(n_layers)])
+        lifting_net = FeedForward(dimension + input_numb_fields, inner_size,
+                                  inner_size, n_layers, func)
+        projection_net = FeedForward(inner_size + dimension, output_numb_fields,
+                                     inner_size, n_layers, func)
+        super().__init__(lifting_net, integral_net, projection_net)
+
+    def forward(self, x):
+        r"""
+        Forward computation for Averaging Neural Operator. It performs a
+        lifting of the input by the ``lifting_net``. Then different layers
+        of Averaging Neural Operator Blocks are applied.
+        Finally the output is projected to the final dimensionality
+        by the ``projecting_net``.
+
+        :param torch.Tensor x: The input tensor for fourier block,
+            depending on ``dimension`` in the initialization. It expects
+            a tensor :math:`B \times N \times D`,
+            where :math:`B` is the batch_size, :math:`N` the number of points
+            in the mesh, :math:`D` the dimension of the problem, i.e. the sum
+            of ``len(coordinates_indices)+len(field_indices)``.
+        :return: The output tensor obtained from Average Neural Operator.
+        :rtype: torch.Tensor
+        """
+        points_tmp = x.extract(self.coordinates_indices)
+        features_tmp = x.extract(self.field_indices)
+        new_batch = concatenate((features_tmp, points_tmp), dim=2)
+        new_batch = self._lifting_operator(new_batch)
+        new_batch = self._integral_kernels(new_batch)
+        new_batch = concatenate((new_batch, points_tmp), dim=2)
+        new_batch = self._projection_operator(new_batch)
+        return new_batch

pina/model/layers/__init__.py

@@ -10,6 +10,7 @@ __all__ = [
     "FourierBlock3D",
     "PODBlock",
     "PeriodicBoundaryEmbedding",
+    "AVNOBlock",
 ]
 
 from .convolution_2d import ContinuousConvBlock
@@ -22,3 +23,4 @@ from .spectral import (
 from .fourier import FourierBlock1D, FourierBlock2D, FourierBlock3D
 from .pod import PODBlock
 from .embedding import PeriodicBoundaryEmbedding
+from .avno_layer import AVNOBlock

pina/model/layers/avno_layer.py

@@ -0,0 +1,67 @@
+""" Module for Averaging Neural Operator Layer class. """
+
+from torch import nn, mean
+from pina.utils import check_consistency
+
+
+class AVNOBlock(nn.Module):
+    r"""
+    The PINA implementation of the inner layer of the Averaging Neural Operator.
+
+    The operator layer performs an affine transformation where the convolution
+    is approximated with a local average. Given the input function
+    :math:`v(x)\in\mathbb{R}^{\rm{emb}}` the layer computes
+    the operator update :math:`K(v)` as:
+
+    .. math::
+        K(v) = \sigma\left(Wv(x) + b + \frac{1}{|\mathcal{A}|}\int v(y)dy\right)
+
+    where:
+
+    *   :math:`\mathbb{R}^{\rm{emb}}` is the embedding (hidden) size
+        corresponding to the ``hidden_size`` object
+    *   :math:`\sigma` is a non-linear activation, corresponding to the
+        ``func`` object
+    *   :math:`W\in\mathbb{R}^{\rm{emb}\times\rm{emb}}` is a tunable matrix.
+    *   :math:`b\in\mathbb{R}^{\rm{emb}}` is a tunable bias.
+
+    .. seealso::
+
+        **Original reference**: Lanthaler S. Li, Z., Kovachki, 
+        Stuart, A. (2020). *The Nonlocal Neural Operator: Universal
+        Approximation*.
+        DOI: `arXiv preprint arXiv:2304.13221.
+        <https://arxiv.org/abs/2304.13221>`_
+
+    """
+
+    def __init__(self, hidden_size=100, func=nn.GELU):
+        """
+        :param int hidden_size: Size of the hidden layer, defaults to 100.
+        :param func: The activation function, default to nn.GELU.
+        """
+        super().__init__()
+
+        # Check type consistency
+        check_consistency(hidden_size, int)
+        check_consistency(func, nn.Module, subclass=True)
+        # Assignment
+        self._nn = nn.Linear(hidden_size, hidden_size)
+        self._func = func()
+
+    def forward(self, x):
+        r"""
+        Forward pass of the layer, it performs a sum of local average
+        and an affine transformation of the field.
+
+        :param torch.Tensor x: The input tensor for performing the
+            computation. It expects a tensor :math:`B \times N \times D`,
+            where :math:`B` is the batch_size, :math:`N` the number of points
+            in the mesh, :math:`D` the dimension of the problem. In particular
+            :math:`D` is the codomain of the function :math:`v`. For example
+            a scalar function has :math:`D=1`, a 4-dimensional vector function
+            :math:`D=4`.
+        :return: The output tensor obtained from Average Neural Operator Block.
+        :rtype: torch.Tensor
+        """
+        return self._func(self._nn(x) + mean(x, dim=1, keepdim=True))

tests/test_model/test_avno.py

@@ -0,0 +1,62 @@
+import torch
+from pina.model import AveragingNeuralOperator
+from pina import LabelTensor
+
+output_numb_fields = 5
+batch_size = 15
+
+
+def test_constructor():
+    input_numb_fields = 1
+    output_numb_fields = 1
+    #minimuum constructor
+    AveragingNeuralOperator(input_numb_fields,
+         output_numb_fields,
+         coordinates_indices=['p'],
+         field_indices=['v'])
+
+    #all constructor
+    AveragingNeuralOperator(input_numb_fields,
+         output_numb_fields,
+         inner_size=5,
+         n_layers=5,
+         func=torch.nn.ReLU,
+         coordinates_indices=['p'],
+         field_indices=['v'])
+
+
+def test_forward():
+    input_numb_fields = 1
+    output_numb_fields = 1
+    dimension = 1
+    input_ = LabelTensor(
+        torch.rand(batch_size, 1000, input_numb_fields + dimension), ['p', 'v'])
+    ano = AveragingNeuralOperator(input_numb_fields,
+               output_numb_fields,
+               dimension=dimension,
+               coordinates_indices=['p'],
+               field_indices=['v'])
+    out = ano(input_)
+    assert out.shape == torch.Size(
+        [batch_size, input_.shape[1], output_numb_fields])
+
+
+def test_backward():
+    input_numb_fields = 1
+    dimension = 1
+    output_numb_fields = 1
+    input_ = LabelTensor(
+        torch.rand(batch_size, 1000, dimension + input_numb_fields), 
+        ['p', 'v'])
+    input_ = input_.requires_grad_()
+    avno = AveragingNeuralOperator(input_numb_fields,
+                output_numb_fields,
+                dimension=dimension,
+                coordinates_indices=['p'],
+                field_indices=['v'])
+    out = avno(input_)
+    tmp = torch.linalg.norm(out)
+    tmp.backward()
+    grad = input_.grad
+    assert grad.shape == torch.Size(
+        [batch_size, input_.shape[1], dimension + input_numb_fields])