.. _ml-ex1: Ex. 1: Surrogate Modeling with a Neural Network ----------------------------------------------- .. note:: This example mirrors the :ref:`sma-ex1` tutorial but instead of using the :ref:`sma-rsm` node, it uses the :ref:`ml-nn_regressor` node, which is the first demonstration case for Nodeworks interfacing with PyTorch. Let's construct a 2D function :math:`r(x,y)` that's: linear in :math:`x`, cubic in :math:`y`, exponentially decays away from the origin and has a little noise to mimic realistic responses of a computational model. In the following example we'll take our "model" to be the simple function defined by: .. math:: r = 2.5 \left(x + y^2 + \mathcal{N}(0,0.1) \right) e^{-(x^2 + y^2)} In the following steps, we'll use nodeworks to sample a relevant :math:`(x,y)` space and construct a surrogate model from responses of the function :math:`r` evaluated at the sample points. Sampling ++++++++ First, we need to statistically design a set of samples to evaluate the model. Launch Nodeworks and add a :ref:`sma-doe` node to the sheet. On the ``Variables`` tab, click the |add| button to create a new variable. Change the name to ``x`` and change the ``from`` and ``to`` values to ``-2.2`` and ``2.2``, respectively. Add another new variable by pressing the |add| button, this time change the name to ``y`` and again change the ``from`` and ``to`` values to ``-2.2`` and ``2.2``, respectively. Click over to the ``Design`` tab, set the ``Method`` to ``factorial`` and the ``Levels`` to ``23``, which will divide the interval between -2.2 and 2.2 into 23 different values of x and y, resulting in 529 total samples. Build the design by clicking on the ``Build`` button to generate the sample. The uniformly spaced factorial design for two input parameters, which is also referred as a 2D factorial, can be viewed on the ``Plot`` tab, as shown below. .. figure:: ./images/ex1_doe.png :align: center Model Evaluation ++++++++++++++++ Now we need to evaluate the model at the sampling points. In this case, we are using the simple function :math:`r`, which can be easily evaluated with a Code node following the steps below: 1. Right click or type to access the node menu and add a ``Code`` node. 2. Set the arguments entry filed at the top of the code node to ``xy`` (since we will pass into the code node an array of the :math:`(x,y)` values at the 529 design points). 3. Evaluate the model function :math:`r` with python. Non-python users can simply copy and paste in the code below into the ``function`` entry. .. code-block:: python import numpy as np returnOut = 2.5*np.exp(-1*(xy[:, 0]**2 + xy[:, 1]**2))*(xy[:, 0] + xy[:, 1]**2 + np.random.normal(0, 0.1, 529)) Then 4. Set the ``Output Selection`` in the DOE node to ``DOE Array`` (our :math:`(x,y)` array). 5. Add a connection from the ``Selected Output`` terminal of the :ref:`sma-doe` to the ``xy`` terminal of the ``Code`` node. .. figure:: ./images/ex1_code.png :align: center Neural Network Training +++++++++++++++++++++++ Now, we will take the function :math:`r` evaluated at the :math:`(x,y)` sample points and train a neural network (NN), to approximate the "model" by following these steps: 1. Right click or type to access the node menu and add a :ref:`ml-nn_regressor` node, found in the ``Machine Learning`` node collection. 2. Connect **both** the ``DOE Matrix`` terminal on the :ref:`sma-doe` node and the ``returnOut`` terminal of the ``code`` node to the ``matrix/response`` terminal of the :ref:`ml-nn_regressor` node. Next we will build the layers of the NN by going to the ``Model`` tab of the :ref:`ml-nn_regressor` node. Click and drag the layers from the ``Available layers`` list to the ``Model layers`` list in the following order: 1. Linear 2. softsign 3. Linear 4. softsign 5. Linear 6. softsign We will leave the options for the layers at their default settings. .. note:: Since the function :math:`r` produces both positive and negative values, the NN needs to end with either an activation function that supports negative values (like ``Tanh`` or ``Softsign``) or a ``Linear`` layer. The shape of the activation function with the default values is displayed under ``layer options``. .. figure:: ./images/ex1_layers.png :align: center Next, on the ``Train`` tab, change the ``Optimizer`` to ``Rprop``, the ``Loss function`` to ``MSE`` and the number of ``Epochs`` to ``200``. Finally, run the sheet by pressing the |play| button. The samples will be evaluated and the NN will be trained. .. figure:: ./images/ex1_train.png :align: center Notice that in the above training plot, once the number of Epochs reaches 100, the error of the test data no longer reduces. This suggests that any further training will result in overfitting or memorization of the training data. On the ``Error`` tab with the Plot type set to "Parity" plot, we can see that the NN has trouble fitting the extremes (the minimum and maximum of the model). This most likely has to do with the low sample density at those locations. When Plot type is set to “errorâ€, the prediction error for each response is plotted. The histogram of errors can be visualized by setting the Plot type to “histogramâ€. Ideally for well fitted model, one would expect to see majority of the errors centered around 0 with low spread in the histogram. .. figure:: ./images/ex1_parity.png :align: center The plot tab enables both 3D and 2D visualization of the constructed NN model for qualitative visualization for the two input parameters while superimposing the actual response values on top of the surface plot. Now you can play with the layers of the model as well as the training properties and the model samples. Can you train a better NN to represent the function :math:`r`? .. |add| image:: ../../../nodeworks/images/add.svg .. |play| image:: ../../../nodeworks/images/play.svg