def tune():
X, y = get_data()
too = torch.optim.Adam, torch.optim.Adadelta, torch.optim.Adagrad, torch.optim.ASGD
to = ht.CategoricalParameter('torch_optimizer', options=too)
eta = ht.ContinuousParameter('eta', lower_bound=1e-10, upper_bound=1e-1)
mi = ht.DiscreteParameter('max_iter', lower_bound=1e2, upper_bound=1e4)
hl1 = ht.DiscreteParameter('', lower_bound=10, upper_bound=100)
hl2 = ht.DiscreteParameter('', lower_bound=10, upper_bound=100)
hls = ht.TupleParameter('hidden_layer_sizes', values=(hl1, hl2))
tp1 = ht.CategoricalParameter('', options=(nn.Linear, ))
tp2 = ht.CategoricalParameter('', options=(nn.Linear, ))
tp3 = ht.CategoricalParameter('', options=(nn.Linear, ))
top = ht.TupleParameter('topology', values=(tp1, tp2, tp3))
hypers = [to, eta, mi, hls, top]
tuner = ht.HyperTune(algorithm=Net,
parameters=hypers,
train_func=Net.fit,
objective_func=Net.mse,
train_func_args=(X, y),
objective_func_args=(X, y),
max_evals=100,
maximize=False,
num_replications=1)
tuner.tune()
print(tuner.get_results())
def test_continuous_param(self):
c = ht.ContinuousParameter('c', lower_bound=0, upper_bound=0)
self.assertEqual(c.shape, 1)
r = rand()
v = c.get_val(r)
self.assertEqual(v, 0)
c = ht.ContinuousParameter('c', lower_bound=0, upper_bound=1)
r = rand()
v = c.get_val(r)
exp = (r[0] + 1) / 2
self.assertEqual(v, exp)
lb, ub = -100, 500
c = ht.ContinuousParameter('c', lower_bound=lb, upper_bound=ub)
r = rand()
v = c.get_val(r)
exp = ((r[0] + 1) / 2) * (ub - lb) + lb
self.assertEqual(v, exp)
def test_categorical_param(self):
c = ht.CategoricalParameter('c', options=('a', 'b'))
self.assertEqual(c.name, 'c')
self.assertEqual(c.shape, 1)
v = c.get_val([-1])
self.assertEqual(v, 'a')
co = ht.ContinuousParameter('co', lower_bound=0, upper_bound=1)
con = ht.ConstantParameter('con', value=100)
c = ht.CategoricalParameter('c', options=('a', co, con))
self.assertEqual(c.shape, 2)
def test_import(self):
print('go')
a = ht.ContinuousParameter('a', lower_bound=0, upper_bound=1)
hypers = [a]
gs = ht.optimizers.GridSearch(depth=1, resolution=0.1)
tuner = ht.HyperTune(algorithm=A,
parameters=hypers,
optimizer=gs,
train_func=A.fit,
objective_func=A.acc,
max_evals=100,
maximize=False,
num_replications=1)
results = tuner.tune()
print(results)
def test_tuple_param(self):
t = ht.TupleParameter('t', values=(1, 1, 1))
self.assertEqual(t.shape, 0)
c = ht.ContinuousParameter('c', lower_bound=0, upper_bound=1)
d = ht.DiscreteParameter('d', lower_bound=0, upper_bound=1)
co = ht.ConstantParameter('co', value='world')
t = ht.TupleParameter('t', values=(c, d, co))
self.assertEqual(t.shape, 2)
r = [-1, 0.5]
v = t.get_val(r)
exp = (0.0, 1, 'world')
self.assertEqual(v, exp)
t = ht.TupleParameter('t', values=(c, {'z': -1}, d))
r = [-1, 0.5]
v = t.get_val(r)
exp = (0.0, {'z': -1}, 1)
self.assertEqual(v, exp)
from sklearn.neural_network import MLPRegressor
import hypertune as ht
import numpy as np
# make an example dataset
p = 75
X = np.random.rand(100, 3)
y = np.random.rand(100)
X_train, X_test, y_train, y_test = X[:p], X[p:], y[:p], y[p:]
# define the target hyperparameters
activation = ht.CategoricalParameter('activation',
options=('identity', 'logistic', 'tanh',
'relu'))
learning_rate_init = ht.ContinuousParameter('learning_rate_init',
lower_bound=10**-5,
upper_bound=0.1)
max_iter = ht.DiscreteParameter('max_iter', lower_bound=500, upper_bound=10**3)
hl1 = ht.DiscreteParameter('', lower_bound=50, upper_bound=250)
hl2 = ht.DiscreteParameter('', lower_bound=100, upper_bound=250)
hl3 = ht.DiscreteParameter('', lower_bound=1, upper_bound=100)
hidden_layer_sizes = ht.TupleParameter('hidden_layer_sizes',
values=(hl1, hl2, hl3))
hypers = [
activation, learning_rate_init, max_iter, learning_rate, hidden_layer_sizes
]
# define a Hypertune object
tuner = ht.HyperTune(algorithm=MLPRegressor,
def mse(self, X, y):
y_hat = self.predict(X)
return self._loss_func_(y, y_hat).item()
# make an example dataset
p = 750
X = np.random.rand(10**4, 3)
y = np.random.rand(10**4)
X_train, X_test, y_train, y_test = X[:p], X[p:], y[:p], y[p:]
# define the target hyperparameters
too = torch.optim.Adam, torch.optim.Adadelta, torch.optim.Adagrad, torch.optim.ASGD
to = ht.CategoricalParameter('torch_optimizer', options=too)
eta = ht.ContinuousParameter('eta', lower_bound=1e-10, upper_bound=1e-1)
mi = ht.DiscreteParameter('max_iter', lower_bound=1e2, upper_bound=1e4)
hl1 = ht.DiscreteParameter('', lower_bound=10, upper_bound=100)
hl2 = ht.DiscreteParameter('', lower_bound=10, upper_bound=100)
hls = ht.TupleParameter('hidden_layer_sizes', values=(hl1, hl2))
tp1 = ht.CategoricalParameter('', options=(nn.Linear, ))
tp2 = ht.CategoricalParameter('', options=(nn.Linear, ))
tp3 = ht.CategoricalParameter('', options=(nn.Linear, ))
top = ht.TupleParameter('topology', values=(tp1, tp2, tp3))
hypers = [to, eta, mi, hls, top]
# define a Hypertune object
tuner = ht.HyperTune(algorithm=Net,
parameters=hypers,
train_func=Net.fit,
objective_func=Net.mse,
return mse
def aux_objective_func(algo, X, y):
_y = np.where(y >= 0.0, 1, -1)
y_hat = np.where(algo.predict(X) >= 0.0, 1, -1)
return np.mean(_y == y_hat)
X, y = datasets.iris(return_splits=False)
X = X[y != 2]
y = y[y != 2]
y[y == 0] = -1
X_train, X_test, y_train, y_test = datasets.split(X, y)
alpha = ht.ContinuousParameter('alpha', lower_bound=10**-10, upper_bound=0.01)
epochs = ht.DiscreteParameter('epochs', lower_bound=200, upper_bound=10**4)
hypers = [alpha, epochs]
tuner = ht.HyperTune(algorithm=Perceptron,
parameters=hypers,
train_func=Perceptron.fit,
objective_func=Perceptron.mse,
train_func_args=(X_train, y_train),
objective_func_args=(X_test, y_test),
max_evals=50,
maximize=False)
results = tuner.tune()
print(results)