def runTest(self):
from dcgpy import symbolic_regression, generate_koza_quintic, kernel_set_double
import pygmo as pg
X, Y = generate_koza_quintic()
udp = symbolic_regression(points=X,
labels=Y,
rows=1,
cols=20,
levels_back=21,
arity=2,
kernels=kernel_set_double(["sum", "diff"])(),
n_eph=2,
multi_objective=False,
parallel_batches=0)
prob = pg.problem(udp)
pop = pg.population(prob, 10)
udp.pretty(pop.champion_x)
udp.prettier(pop.champion_x)
# Unconstrained
self.assertEqual(prob.get_nc(), 0)
self.assertEqual(prob.get_nic(), 0)
# Single objective
self.assertEqual(prob.get_nobj(), 1)
# Dimensions
self.assertEqual(prob.get_nix(), 20 * (2 + 1) + 1)
self.assertEqual(prob.get_nx(), 2 + prob.get_nix())
# Has gradient and hessians
self.assertEqual(prob.has_gradient(), True)
self.assertEqual(prob.has_hessians(), True)
def runTest(self):
from dcgpy import symbolic_regression, generate_koza_quintic, kernel_set_double, gd4cgp
import pygmo as pg
import pickle
X, Y = generate_koza_quintic()
# Interface for the UDPs
udp = symbolic_regression(points=X,
labels=Y,
rows=1,
cols=20,
levels_back=21,
arity=2,
kernels=kernel_set_double(
["sum", "diff", "mul", "pdiv"])(),
n_eph=2,
multi_objective=False,
parallel_batches=0)
prob = pg.problem(udp)
pop = pg.population(prob, 10)
# Interface for the UDAs
uda = gd4cgp(max_iter=10, lr=0.1, lr_min=1e-6)
algo = pg.algorithm(uda)
algo.set_verbosity(0)
# Testing some evolutions
pop = algo.evolve(pop)
# In parallel
archi = pg.archipelago(prob=prob, algo=algo, n=16, pop_size=4)
archi.evolve()
archi.wait_check()
# Pickling.
self.assertTrue(repr(algo) == repr(pickle.loads(pickle.dumps(algo))))
def run(dataset_train, dataset_test, cols, gen):
ss = dcgpy.kernel_set_double([
"sum", "diff", "mul", "pdiv", "sin", "cos", "tanh", "log", "exp",
"psqrt"
])
Xtrain, ytrain = dataset_train[:, :-1], dataset_train[:, -1]
Xtest, ytest = dataset_test[:, :-1], dataset_test[:, -1]
udp = dcgpy.symbolic_regression(points=Xtrain,
labels=ytrain[:, np.newaxis],
kernels=ss(),
rows=1,
cols=cols,
levels_back=21,
arity=2,
n_eph=3,
multi_objective=False,
parallel_batches=0)
uda = dcgpy.es4cgp(gen=gen) #, mut_n = 1)
algo = pg.algorithm(uda)
pop = pg.population(udp, 4)
pop = algo.evolve(pop)
return RMSE(udp.predict(Xtrain, pop.champion_x),
ytrain), RMSE(udp.predict(Xtest, pop.champion_x), ytest)
def encode_ffnn(inputs, outputs, layers_size, kernels, levels_back):
"""
encode_ffnn(inputs, outputs, layers, kernels, levels_back)
Encodes a feed forward neural network as a dCGPANN expression. While there are infinite ways to perform such an encoding
this function generates one of the possible ones.
Args:
inputs (``int``): number of inputs
outputs (``int``): number of outputs
layers_size (``List[int]``): size of hidden neural layers
kernels (``List[string])``: list containing the non linearity name for each hidden neural layer plus the one for the output layer
levels_back (``int``): number of levels-back in the cartesian program
Returns:
A ``expression_ann_double`` encoding the requested network.
Raises:
ValueError: if the kernel list contains unknown kernel names or if layers_size and kernels are malformed
"""
from dcgpy import expression_ann_double, kernel_set_double
if (len(kernels) != len(layers_size) + 1):
raise ValueError(
"The size of layers_size must be one less than the size of kernels (as this also includes the output nonlinearity) "
)
# We create the list of possible kernels
kernel_list = kernel_set_double(list(set(kernels)))()
# We compute the cartesian substrate able to host the network
arity = [inputs] + layers_size
extended_layer_size = layers_size + [outputs]
cols = len(extended_layer_size)
rows = max(extended_layer_size)
# We create a dCGPANN expression that is able to encode the FFNN
retval = expression_ann_double(inputs, outputs, rows, cols, levels_back,
arity, kernel_list)
# We hand write the chromosome.
start_prev_col = [0] + [inputs + c * rows for c in range(cols)]
x = retval.get()
for c in range(cols):
kernel_id = list(set(kernels)).index(kernels[c])
for r in range(extended_layer_size[c]):
node_id = inputs + c * rows + r
g_idx = retval.get_gene_idx()[node_id]
x[g_idx] = kernel_id
for conn in range(arity[c]):
x[g_idx + conn + 1] = start_prev_col[c] + conn
# And the output
for i in range(outputs):
x[-outputs + i] = start_prev_col[c + 1] + i
retval.set(x)
return retval
def main():
# Some necessary imports.
import dcgpy
import pygmo as pg
# Sympy is nice to have for basic symbolic manipulation.
from sympy import init_printing
from sympy.parsing.sympy_parser import parse_expr
init_printing()
# Fundamental for plotting.
from matplotlib import pyplot as plt
# We load our data from some available ones shipped with dcgpy.
# In this particular case we use the problem chwirut2 from
# (https://www.itl.nist.gov/div898/strd/nls/data/chwirut2.shtml)
X, Y = dcgpy.generate_chwirut2()
# And we plot them as to visualize the problem.
_ = plt.plot(X, Y, '.')
_ = plt.title('54 measurments')
_ = plt.xlabel('metal distance')
_ = plt.ylabel('ultrasonic response')
# We define our kernel set, that is the mathematical operators we will
# want our final model to possibly contain. What to choose in here is left
# to the competence and knowledge of the user. A list of kernels shipped with dcgpy
# can be found on the online docs. The user can also define its own kernels (see the corresponding tutorial).
ss = dcgpy.kernel_set_double(["sum", "diff", "mul", "pdiv"])
# We instantiate the symbolic regression optimization problem (note: many important options are here not
# specified and thus set to their default values)
udp = dcgpy.symbolic_regression(points=X, labels=Y, kernels=ss())
print(udp)
# We instantiate here the evolutionary strategy we want to use to search for models.
uda = dcgpy.es4cgp(gen=10000, max_mut=2)
prob = pg.problem(udp)
algo = pg.algorithm(uda)
# Note that the screen output will happen on the terminal, not on your Jupyter notebook.
# It can be recovered afterwards from the log.
algo.set_verbosity(10)
pop = pg.population(prob, 4)
pop = algo.evolve(pop)
# Lets have a look to the symbolic representation of our model (using sympy)
parse_expr(udp.prettier(pop.champion_x))
# And lets see what our model actually predicts on the inputs
Y_pred = udp.predict(X, pop.champion_x)
# Lets comapre to the data
_ = plt.plot(X, Y_pred, 'r.')
_ = plt.plot(X, Y, '.', alpha=0.2)
_ = plt.title('54 measurments')
_ = plt.xlabel('metal distance')
_ = plt.ylabel('ultrasonic response')
# Here we get the log of the latest call to the evolve
log = algo.extract(dcgpy.es4cgp).get_log()
gen = [it[0] for it in log]
loss = [it[2] for it in log]
# And here we plot, for example, the generations against the best loss
_ = plt.semilogy(gen, loss)
_ = plt.title('last call to evolve')
_ = plt.xlabel('metal distance')
_ = plt.ylabel('ultrasonic response')
# We load our data from some available ones shipped with dcgpy.
# In this particular case we use the problem chwirut2 from
# (https://www.itl.nist.gov/div898/strd/nls/data/chwirut2.shtml)
X, Y = dcgpy.generate_chwirut2()
# And we plot them as to visualize the problem.
_ = plt.plot(X, Y, '.')
_ = plt.title('54 measurments')
_ = plt.xlabel('metal distance')
_ = plt.ylabel('ultrasonic response')
# We define our kernel set, that is the mathematical operators we will
# want our final model to possibly contain. What to choose in here is left
# to the competence and knowledge of the user. A list of kernels shipped with dcgpy
# can be found on the online docs. The user can also define its own kernels (see the corresponding tutorial).
ss = dcgpy.kernel_set_double(["sum", "diff", "mul", "pdiv"])
# We instantiate the symbolic regression optimization problem (note: many important options are here not
# specified and thus set to their default values)
udp = dcgpy.symbolic_regression(points = X, labels = Y, kernels=ss())
print(udp)
# We instantiate here the evolutionary strategy we want to use to search for models.
uda = dcgpy.es4cgp(gen = 10000, mut_n = 2)
prob = pg.problem(udp)
algo = pg.algorithm(uda)
# Note that the screen output will happen on the terminal, not on your Jupyter notebook.
# It can be recovered afterwards from the log.
algo.set_verbosity(10)