def run_demo_1D(is_gradient_enhancement=True): # pragma: no cover
"""Test and demonstrate GENN using a 1D example"""
# Test function
f = lambda x: x * np.sin(x)
df_dx = lambda x: np.sin(x) + x * np.cos(x)
# Domain
lb = -np.pi
ub = np.pi
# Training data
m = 4
xt = np.linspace(lb, ub, m)
yt = f(xt)
dyt_dxt = df_dx(xt)
# Validation data
xv = lb + np.random.rand(30, 1) * (ub - lb)
yv = f(xv)
dyv_dxv = df_dx(xv)
# Initialize GENN object
genn = GENN()
genn.options["alpha"] = 0.05
genn.options["beta1"] = 0.9
genn.options["beta2"] = 0.99
genn.options["lambd"] = 0.05
genn.options["gamma"] = int(is_gradient_enhancement)
genn.options["deep"] = 2
genn.options["wide"] = 6
genn.options["mini_batch_size"] = 64
genn.options["num_epochs"] = 25
genn.options["num_iterations"] = 100
genn.options["seed"] = SEED
genn.options["is_print"] = True
# Load data
load_smt_data(genn, xt, yt, dyt_dxt)
# Train
genn.train()
genn.plot_training_history()
genn.goodness_of_fit(xv, yv, dyv_dxv)
# Plot comparison
if genn.options["gamma"] == 1.0:
title = 'with gradient enhancement'
else:
title = 'without gradient enhancement'
x = np.arange(lb, ub, 0.01)
y = f(x)
y_pred = genn.predict_values(x)
fig, ax = plt.subplots()
ax.plot(x, y_pred)
ax.plot(x, y, 'k--')
ax.plot(xv, yv, 'ro')
ax.plot(xt, yt, 'k+', mew=3, ms=10)
ax.set(xlabel='x', ylabel='y', title=title)
ax.legend(['Predicted', 'True', 'Test', 'Train'])
plt.show()
def test_genn(self):
import numpy as np
import matplotlib.pyplot as plt
from smt.surrogate_models.genn import GENN, load_smt_data
# Training data
lower_bound = -np.pi
upper_bound = np.pi
number_of_training_points = 4
xt = np.linspace(lower_bound, upper_bound, number_of_training_points)
yt = xt * np.sin(xt)
dyt_dxt = np.sin(xt) + xt * np.cos(xt)
# Validation data
number_of_validation_points = 30
xv = np.linspace(lower_bound, upper_bound, number_of_validation_points)
yv = xv * np.sin(xv)
dyv_dxv = np.sin(xv) + xv * np.cos(xv)
# Truth model
x = np.arange(lower_bound, upper_bound, 0.01)
y = x * np.sin(x)
# GENN
genn = GENN()
genn.options[
"alpha"] = 0.1 # learning rate that controls optimizer step size
genn.options[
"beta1"] = 0.9 # tuning parameter to control ADAM optimization
genn.options[
"beta2"] = 0.99 # tuning parameter to control ADAM optimization
genn.options[
"lambd"] = 0.1 # lambd = 0. = no regularization, lambd > 0 = regularization
genn.options[
"gamma"] = 1.0 # gamma = 0. = no grad-enhancement, gamma > 0 = grad-enhancement
genn.options["deep"] = 2 # number of hidden layers
genn.options["wide"] = 6 # number of nodes per hidden layer
genn.options[
"mini_batch_size"] = 64 # used to divide data into training batches (use for large data sets)
genn.options["num_epochs"] = 20 # number of passes through data
genn.options[
"num_iterations"] = 100 # number of optimizer iterations per mini-batch
genn.options["is_print"] = True # print output (or not)
load_smt_data(
genn, xt, yt, dyt_dxt
) # convenience function to read in data that is in SMT format
genn.train() # API function to train model
genn.plot_training_history(
) # non-API function to plot training history (to check convergence)
genn.goodness_of_fit(
xv, yv,
dyv_dxv) # non-API function to check accuracy of regression
y_pred = genn.predict_values(
x) # API function to predict values at new (unseen) points
# Plot
fig, ax = plt.subplots()
ax.plot(x, y_pred)
ax.plot(x, y, "k--")
ax.plot(xv, yv, "ro")
ax.plot(xt, yt, "k+", mew=3, ms=10)
ax.set(xlabel="x", ylabel="y", title="GENN")
ax.legend(["Predicted", "True", "Test", "Train"])
plt.show()
def run_demo_2d(alpha=0.1,
beta1=0.9,
beta2=0.99,
lambd=0.1,
gamma=1.0,
deep=3,
wide=6,
mini_batch_size=None,
iterations=30,
epochs=100):
"""
Predict Rastrigin function using neural net and compare against truth model. Provided with proper training data,
the only hyperparameters the user needs to tune are:
:param alpha = learning rate
:param beta1 = adam optimizer parameter
:param beta2 = adam optimizer parameter
:param lambd = regularization coefficient
:param gamma = gradient enhancement coefficient
:param deep = neural net depth
:param wide = neural net width
This restricted list is intentional. The goal was to provide a simple interface for common regression tasks
with the bare necessary tuning parameters. More advanced prediction tasks should consider tensorflow or other
deep learning frameworks. Hopefully, the simplicity of this interface will address a common use case in aerospace
engineering, namely: predicting smooth functions using computational design of experiments.
"""
if gamma > 0.:
title = 'GENN'
else:
title = 'NN'
# Practice data
X_train, Y_train, J_train = get_practice_data(random=False)
X_test, Y_test, J_test = get_practice_data(random=True)
# Convert training data to SMT format
xt = X_train.T
yt = Y_train.T
dyt_dxt = J_train[
0].T # SMT format doesn't handle more than one output at a time, hence J[0]
# Convert test data to SMT format
xv = X_test.T
yv = Y_test.T
dyv_dxv = J_test[
0].T # SMT format doesn't handle more than one output at a time, hence J[0]
# Initialize GENN object
genn = GENN()
genn.options["alpha"] = alpha
genn.options["beta1"] = beta1
genn.options["beta2"] = beta2
genn.options["lambd"] = lambd
genn.options["gamma"] = gamma
genn.options["deep"] = deep
genn.options["wide"] = wide
genn.options["mini_batch_size"] = mini_batch_size
genn.options["num_epochs"] = epochs
genn.options["num_iterations"] = iterations
genn.options["seed"] = SEED
genn.options["is_print"] = True
# Load data
load_smt_data(
genn, xt, yt, dyt_dxt
) # convenience function that uses SurrogateModel.set_training_values(), etc.
# Train
genn.train()
genn.plot_training_history()
genn.goodness_of_fit(xv, yv, dyv_dxv)
# Contour plot
contour_plot(genn, title=title)
