From 9c5d331f22236b8da3f4379fa2ec87d30ac9b50d Mon Sep 17 00:00:00 2001
From: Jonathan Lee <nathanee@seas.upenn.edu>
Date: Tue, 15 Nov 2022 20:45:52 -0500
Subject: [PATCH] Add documentation for Function Approximation from a Formula
 (#1034)

---
 docs/demos/function.rst      |  6 ++-
 docs/demos/function/func.rst | 83 ++++++++++++++++++++++++++++++++++++
 2 files changed, 88 insertions(+), 1 deletion(-)
 create mode 100644 docs/demos/function/func.rst

diff --git a/docs/demos/function.rst b/docs/demos/function.rst
index 0dadc9274..d8d0785d1 100644
--- a/docs/demos/function.rst
+++ b/docs/demos/function.rst
@@ -6,8 +6,12 @@ Here are some demos of learning functions.
 Function approximation
 ----------------------
 
+.. toctree::
+   :maxdepth: 1
+
+   function/func
+
 - `Learning a function from a dataset <https://github.com/lululxvi/deepxde/blob/master/examples/function/dataset.py>`_
-- `Learning a function from the formula <https://github.com/lululxvi/deepxde/blob/master/examples/function/func.py>`_
 
 Uncertainty quantification
 --------------------------
diff --git a/docs/demos/function/func.rst b/docs/demos/function/func.rst
new file mode 100644
index 000000000..852a6a292
--- /dev/null
+++ b/docs/demos/function/func.rst
@@ -0,0 +1,83 @@
+Learning a function from a formula
+===================
+
+Problem setup
+-------------
+
+We will solve a simple function approximation problem from a formula:
+
+.. math:: f(x) = x * \sin(5x)
+
+Implementation
+--------------
+
+This description goes through the implementation of a solver for the above function step-by-step.
+
+First, the DeepXDE and NumPy (``np``) modules are imported:
+
+.. code-block:: python
+
+    import deepxde as dde
+    import numpy as np
+
+We begin by defining a simple function which will be approximated. 
+
+.. code-block:: python
+
+    def func(x):
+        """
+        x: array_like, N x D_in
+        y: array_like, N x D_out
+        """
+        return x * np.sin(5 * x)
+
+The argument ``x`` to ``func`` is the network input. The ``func`` simply returns the corresponding function values from the given ``x``. 
+
+Then, we define a computational domain. We can use a built-in class ``Interval`` as follows:
+
+.. code-block:: python
+
+    geom = dde.geometry.Interval(-1, 1)
+
+Now, we have specified the geometry, we need to define the problem using a built-in class ``Function``
+
+.. code-block:: python
+
+    num_train = 16
+    num_test = 100
+    data = dde.data.Function(geom, func, num_train, num_test)
+
+Here, we use 16 points for training sampled inside the domain, and 100 points for testing.
+
+Next, we choose a fully connected neural network of depth 4 (i.e., 3 hidden layers) and width 20 with ``tanh`` as the activation function and ``Glorot uniform`` as the initializer:
+
+.. code-block:: python
+
+    activation = "tanh"
+    initializer = "Glorot uniform"
+    net = dde.nn.FNN([1] + [20] * 3 + [1], activation, initializer)
+
+Now, we have the function approximation problem and the network. We bulid a ``Model`` and choose the optimizer ``adam`` and the learning rate of ``0.001``:
+
+.. code-block:: python
+
+    model = dde.Model(data, net)
+    model.compile("adam", lr=0.001, metrics=["l2 relative error"])
+   
+We then train the model for 10000 iterations:
+
+.. code-block:: python
+
+    losshistory, train_state = model.train(iterations=10000)
+    
+We also save and plot the best trained result and loss history.
+
+.. code-block:: python
+
+    dde.saveplot(losshistory, train_state, issave=True, isplot=True)
+
+Complete code
+-------------
+
+.. literalinclude:: ../../../examples/function/func.py
+  :language: python