def objective(individual: IndividualSingleGenome, seed, n_runs_per_individual):
print(individual.to_sympy())
f = individual.to_numpy()
# with open("callable.pkl", "wb") as fi:
# pickle.dump(f, fi)
individual.fitness = inner_objective(f, seed, n_runs_per_individual)
return individual
def test_pickle_individual():
primitives = (cgp.Add,)
genome = cgp.Genome(1, 1, 1, 1, 1, primitives)
individual = IndividualSingleGenome(None, genome)
with open("individual.pkl", "wb") as f:
pickle.dump(individual, f)
def test_update_parameters_from_torch_class_does_not_reset_fitness_for_unused_parameters():
pytest.importorskip("torch")
primitives = (cgp.Mul, cgp.Parameter)
genome = cgp.Genome(1, 1, 2, 1, 1, primitives)
# f(x) = x ** 2
genome.dna = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
1, # these
0, # three
0, # genes code for an unused parameter node
0,
0,
0,
ID_OUTPUT_NODE,
2,
ID_NON_CODING_GENE,
]
fitness = 1.0
individual = IndividualSingleGenome(fitness, genome)
f = individual.to_torch()
assert individual.fitness is not None
individual.update_parameters_from_torch_class(f)
assert individual.fitness is not None
g = individual.to_func()
x = 2.0
assert g([x])[0] == pytest.approx(x ** 2)
def test_update_parameters_from_torch_class_resets_fitness():
pytest.importorskip("torch")
primitives = (cgp.Mul, cgp.Parameter)
genome = cgp.Genome(1, 1, 2, 1, 1, primitives)
# f(x) = c * x
genome.dna = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
1,
0,
0,
0,
0,
1,
ID_OUTPUT_NODE,
2,
ID_NON_CODING_GENE,
]
fitness = 1.0
individual = IndividualSingleGenome(fitness, genome)
f = individual.to_torch()
f._p1.data[0] = math.pi
assert individual.fitness is not None
individual.update_parameters_from_torch_class(f)
assert individual.fitness is None
g = individual.to_func()
x = 2.0
assert g([x])[0] == pytest.approx(math.pi * x)
def test_individual_with_parameter_torch():
torch = pytest.importorskip("torch")
primitives = (cgp.Add, cgp.Parameter)
genome = cgp.Genome(1, 1, 2, 1, 2, primitives)
# f(x) = x + c
genome.dna = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
1,
0,
0,
0,
0,
1,
ID_OUTPUT_NODE,
2,
ID_NON_CODING_GENE,
]
individual = IndividualSingleGenome(None, genome)
c = 1.0
x = torch.empty(2, 1).normal_()
f = individual.to_torch()
y = f(x)
assert y[0, 0].item() == pytest.approx(x[0, 0].item() + c)
assert y[1, 0].item() == pytest.approx(x[1, 0].item() + c)
c = 2.0
individual.genome.parameter_names_to_values["<p1>"] = c
f = individual.to_torch()
y = f(x)
assert y[0, 0].item() == pytest.approx(x[0, 0].item() + c)
assert y[1, 0].item() == pytest.approx(x[1, 0].item() + c)
def test_individual_with_parameter_sympy():
sympy = pytest.importorskip("sympy") # noqa
primitives = (cgp.Add, cgp.Parameter)
genome = cgp.Genome(1, 1, 2, 1, 2, primitives)
# f(x) = x + c
genome.dna = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
1,
0,
0,
0,
0,
1,
ID_OUTPUT_NODE,
2,
ID_NON_CODING_GENE,
]
individual = IndividualSingleGenome(None, genome)
c = 1.0
x = [3.0]
f = individual.to_sympy()[0]
y = f.subs("x_0", x[0]).evalf()
assert y == pytest.approx(x[0] + c)
c = 2.0
individual.genome.parameter_names_to_values["<p1>"] = c
f = individual.to_sympy()[0]
y = f.subs("x_0", x[0]).evalf()
assert y == pytest.approx(x[0] + c)
def test_individual_with_parameter_python():
primitives = (cgp.Add, cgp.Parameter)
genome = cgp.Genome(1, 1, 2, 1, 2, primitives)
# f(x) = x + c
genome.dna = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
1,
0,
0,
0,
0,
1,
ID_OUTPUT_NODE,
2,
ID_NON_CODING_GENE,
]
individual = IndividualSingleGenome(None, genome)
c = 1.0
x = [3.0]
f = individual.to_func()
y = f(x)
assert y[0] == pytest.approx(x[0] + c)
c = 2.0
individual.genome.parameter_names_to_values["<p1>"] = c
f = individual.to_func()
y = f(x)
assert y[0] == pytest.approx(x[0] + c)
def test_to_and_from_torch_plus_backprop():
torch = pytest.importorskip("torch")
primitives = (cgp.Mul, cgp.Parameter)
genome = cgp.Genome(1, 1, 2, 2, 1, primitives)
# f(x) = c * x
genome.dna = [
ID_INPUT_NODE,
ID_NON_CODING_GENE,
ID_NON_CODING_GENE,
1,
0,
0,
1,
0,
0,
0,
0,
1,
0,
0,
1,
ID_OUTPUT_NODE,
3,
ID_NON_CODING_GENE,
]
individual = IndividualSingleGenome(None, genome)
def f_target(x):
return math.pi * x
f = individual.to_torch()
optimizer = torch.optim.SGD(f.parameters(), lr=1e-1)
criterion = torch.nn.MSELoss()
for i in range(200):
x = torch.DoubleTensor(1, 1).normal_()
y = f(x)
y_target = f_target(x)
loss = criterion(y, y_target)
f.zero_grad()
loss.backward()
optimizer.step()
assert loss.detach().numpy() < 1e-15
# use old parameter values to compile function
x = [3.0]
f_func = individual.to_func()
y = f_func(x)
assert y[0] != pytest.approx(f_target(x[0]))
# update parameter values from torch class and compile new
# function with new parameter values
individual.update_parameters_from_torch_class(f)
f_func = individual.to_func()
y = f_func(x)
assert y[0] == pytest.approx(f_target(x[0]))