def boston():
config = dict(
dataset=ShuffledDataset(
data_path='data/boston_housing_data.csv',
shuffle_path='data/boston_housing_data.npy',
standardize=True,
shuffle=0 # shuffle_id
),
split=dict(train_size=0.9, ),
method=PiEnsemble,
hyper_params=dict(ensemble_size=1,
aggreg_func=no_aggreg,
hidden_size=[50, 50],
epochs=5000,
batch_size=100,
optimizer=Adam,
learning_rate=lambda t: t.suggest_discrete_uniform(
'learning_rate', 0.001, 0.01, 0.001),
scheduler=ExponentialDecay,
decay_rate=lambda t: t.suggest_discrete_uniform(
'decay_rate', 0.95, 1., 0.01),
decay_steps=50.,
early_stopping=True,
punish_crossing=False,
patience=500,
delta=1e-6,
tolerance=0.01,
loss_func=mse_loss,
alpha=None,
print_frequency=10,
metrics=[mse],
device='cpu'),
num_trials=300)
def boston():
config = dict(
dataset=ShuffledDataset(
data_path='data/boston_housing_data.csv',
shuffle_path='data/boston_housing_data.npy',
standardize=True,
shuffle=True
),
split=dict(
train_size=0.9,
test_size=0.1
),
num_runs=20,
method=PiEnsemble,
hyper_params=dict( # HPS trial number 227
ensemble_size=5,
aggreg_func=mean_aggreg,
hidden_size=[50, 50],
epochs=129,
batch_size=100,
optimizer=Adam,
learning_rate=0.004,
scheduler=ExponentialDecay,
decay_rate=0.99,
decay_steps=50.,
loss_func=mse_loss,
alpha=None,
metrics=[mse],
print_frequency=10,
device='cpu'
)
)
def protein():
config = dict(
dataset=ShuffledDataset(
data_path='data/protein.csv',
shuffle_path='data/protein.npy',
standardize=True,
shuffle=0 # shuffle_id
),
split=dict(train_size=0.9, ),
method=PiEnsemble,
hyper_params=dict(ensemble_size=1,
aggreg_func=no_aggreg,
hidden_size=[100, 100],
epochs=5000,
batch_size=100,
optimizer=Adam,
learning_rate=lambda t: t.suggest_discrete_uniform(
'learning_rate', 0.001, 0.01, 0.001),
scheduler=ExponentialDecay,
decay_rate=lambda t: t.suggest_discrete_uniform(
'decay_rate', 0.95, 1., 0.01),
decay_steps=50.,
early_stopping=True,
patience=500,
delta=1e-6,
tolerance=0.01,
loss_func=qd_code_loss,
alpha=0.05,
soften=160.,
lambda_=lambda t: t.suggest_discrete_uniform(
'lambda_', 1., 50., 1.),
print_frequency=10,
device='cpu'),
num_trials=300)
def yacht_code():
"""
Original HPs with `alpha=0.01`.
Therefore new HPs with random search with the objective to optimize
for the aggregation as did Pearce et al. (2018);
confirmed in our email communication.
"""
config = dict(dataset=ShuffledDataset(data_path='data/yacht.csv',
shuffle_path='data/yacht.npy',
standardize=True,
shuffle=True),
split=dict(train_size=0.9, test_size=0.1),
num_runs=20,
method=PiEnsemble,
hyper_params=dict(ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg],
hidden_size=[50],
optimizer=Adam,
learning_rate=0.009,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=0.98,
epochs=2500,
batch_size=100,
loss_func=qd_code_loss,
alpha=0.05,
soften=160.,
lambda_=16.,
print_frequency=10))
def boston_dev():
config = dict(dataset=ShuffledDataset(
data_path='data/boston_housing_data.csv',
shuffle_path='data/boston_housing_data.npy',
standardize=True,
shuffle=True),
split=dict(train_size=0.9, test_size=0.1),
num_runs=5,
method=PiEnsemble,
hyper_params=dict(ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg],
hidden_size=[50],
epochs=5,
batch_size=100,
optimizer=Adam,
learning_rate=0.02,
scheduler=ExponentialDecay,
decay_rate=0.9,
decay_steps=50.,
loss_func=qd_code_loss,
alpha=0.05,
soften=160.,
lambda_=15.,
print_frequency=10))
seed = 1
def concrete():
config = dict(
dataset=ShuffledDataset(data_path='data/concrete_data.csv',
shuffle_path='data/concrete_data.npy',
standardize=True,
shuffle=True),
split=dict(
train_size=0.9,
val_size=0.,
test_size=0.1,
),
num_runs=20,
method=PiEnsemble,
hyper_params=dict( # HPS trial number 210
ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg, snm_aggreg],
hidden_size=[50, 50],
optimizer=Adam,
learning_rate=0.007,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=1.,
epochs=680,
batch_size=100,
loss_func=qd_plus_loss,
alpha=0.05,
soften=160.,
lambda_1=0.9958,
lambda_2=0.34,
ksi=10.,
print_frequency=10,
device='cpu'))
def yacht():
config = dict(
dataset=ShuffledDataset(
data_path='data/yacht.csv',
shuffle_path='data/yacht.npy',
standardize=True,
shuffle=0 # shuffle_id
),
split=dict(train_size=0.9, ),
method=MvEnsemble,
hyper_params=dict(ensemble_size=1,
aggreg_func=no_aggreg,
hidden_size=[50, 50],
epochs=1000,
batch_size=100,
optimizer=Adam,
learning_rate=lambda t: t.suggest_discrete_uniform(
'learning_rate', 0.001, 0.01, 0.001),
scheduler=ExponentialDecay,
decay_rate=lambda t: t.suggest_discrete_uniform(
'decay_rate', 0.95, 1., 0.01),
decay_steps=50,
loss_func=normal_loss,
epsilon=None,
alpha=0.05,
early_stopping=True,
patience=100,
delta=1e-6,
tolerance=0.01,
print_frequency=10,
device='cpu'),
num_trials=100)
def protein():
config = dict(
dataset=ShuffledDataset(data_path='data/protein.csv',
shuffle_path='data/protein.npy',
standardize=True,
shuffle=True),
split=dict(train_size=0.9, test_size=0.1),
num_runs=5,
method=PiEnsemble,
hyper_params=dict( # HPS trial number 24 and then manually fine-tuned
ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg, snm_aggreg],
hidden_size=[100, 100],
optimizer=Adam,
learning_rate=0.02,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=0.99,
epochs=80,
batch_size=100,
loss_func=qd_plus_loss,
alpha=0.05,
soften=160.,
lambda_1=0.9976,
lambda_2=0.24,
ksi=10.,
print_frequency=10,
device='cpu'))
def naval():
config = dict(
dataset=ShuffledDataset(data_path='data/naval_compressor_decay.csv',
shuffle_path='data/naval_compressor_decay.npy',
standardize=True,
shuffle=True),
split=dict(train_size=0.9, test_size=0.1),
num_runs=20,
method=PiEnsemble,
hyper_params=dict( # HPS trial number 21 and then manually fine-tuned
ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg, snm_aggreg],
hidden_size=[50, 50],
optimizer=Adam,
learning_rate=0.001,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=0.9975,
epochs=1766,
batch_size=100,
loss_func=qd_plus_loss,
alpha=0.05,
soften=160.,
lambda_1=0.9708,
lambda_2=0.63,
ksi=10.,
print_frequency=10,
device='cpu'))
def boston():
config = dict(
dataset=ShuffledDataset(
data_path='data/boston_housing_data.csv',
shuffle_path='data/boston_housing_data.npy',
standardize=True,
shuffle=True
),
split=dict(
train_size=0.9,
test_size=0.1
),
num_runs=20,
method=MvEnsemble,
hyper_params=dict( # HPS trial number 40
ensemble_size=5,
aggreg_func=[mv_aggreg],
hidden_size=[50, 50],
epochs=152,
batch_size=100,
optimizer=Adam,
learning_rate=0.005,
# scheduler=ExponentialDecay,
# decay_rate=1.,
# decay_steps=50.,
loss_func=normal_loss,
epsilon=0.1, # `None` to disable adversarial examples
alpha=0.05,
print_frequency=10,
device='cpu'
)
)
def yacht():
config = dict(
dataset=ShuffledDataset(data_path='data/yacht.csv',
shuffle_path='data/yacht.npy',
standardize=True,
shuffle=True),
split=dict(
train_size=0.9,
val_size=0.,
test_size=0.1,
),
num_runs=20,
method=PiEnsemble,
hyper_params=dict( # HPS trial number 193
ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg],
hidden_size=[50, 50],
optimizer=Adam,
learning_rate=0.006,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=1.,
epochs=1214,
batch_size=100,
loss_func=qd_code_loss,
alpha=0.05,
soften=160.,
lambda_=1.,
print_frequency=10,
device='cpu'))
def wine():
config = dict(
dataset=ShuffledDataset(
data_path='data/wine.csv',
shuffle_path='data/wine.npy',
standardize=True,
shuffle=0 # shuffle_id
),
split=dict(train_size=0.9, ),
method=PiEnsemble,
hyper_params=dict(
ensemble_size=1,
aggreg_func=no_aggreg,
hidden_size=[50, 50],
epochs=lambda trial: int(
trial.suggest_discrete_uniform('num_epoch', 100, 1000, 40)),
batch_size=100,
optimizer=Adam,
learning_rate=lambda t: t.suggest_discrete_uniform(
'learning_rate', 0.001, 0.01, 0.001),
scheduler=ExponentialDecay,
decay_rate=lambda t: t.suggest_discrete_uniform(
'decay_rate', 0.95, 1., 0.01),
decay_steps=50.,
early_stopping=False,
patience=500,
delta=1e-6,
tolerance=0.01,
loss_func=qd_plus_loss,
alpha=0.05,
soften=160.,
lambda_1=lambda t: t.suggest_discrete_uniform(
'lambda_1', 0.97, 0.999, 0.0002),
lambda_2=lambda t: t.suggest_discrete_uniform(
'lambda_2', 0.01, 0.99, 0.01),
ksi=10.,
print_frequency=10,
device='cpu'),
num_trials=1000)
def yacht_paper():
config = dict(dataset=ShuffledDataset(data_path='data/yacht.csv',
shuffle_path='data/yacht.npy',
standardize=True,
shuffle=True),
split=dict(train_size=0.9, test_size=0.1),
num_runs=20,
method=PiEnsemble,
hyper_params=dict(ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg],
hidden_size=[50],
optimizer=Adam,
learning_rate=0.009,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=0.98,
epochs=2500,
batch_size=100,
loss_func=qd_paper_loss,
alpha=0.05,
soften=160.,
lambda_=16.,
print_frequency=10))
def protein_paper():
config = dict(dataset=ShuffledDataset(data_path='data/protein.csv',
shuffle_path='data/protein.npy',
standardize=True,
shuffle=True),
split=dict(train_size=0.9, test_size=0.1),
num_runs=5,
method=PiEnsemble,
hyper_params=dict(ensemble_size=5,
aggreg_func=[sem_aggreg, std_aggreg],
hidden_size=[100],
optimizer=Adam,
learning_rate=0.002,
scheduler=ExponentialDecay,
decay_steps=50.,
decay_rate=0.999,
epochs=600,
batch_size=100,
loss_func=qd_paper_loss,
retry_on_crossing=False,
alpha=0.05,
soften=160.,
lambda_=40.,
print_frequency=10))
