def __init__(self,
method,
hyper_params,
inputs_block,
attr_types,
random=False,
random_max_run=1,
name="DN"):
assert 'r' not in attr_types, "real value attributes are not allowd here"
self.attr_types = attr_types
self.num_attr = len(self.attr_types)
self.inputs_block = inputs_block
self.method = method
self.num_inputs = self.inputs_block[-1][1]
self.name = name
self.models = []
for i, block in enumerate(self.inputs_block):
self.models.append(self.method())
try:
self.models[-1].set_params(probability=True)
except:
pass
self.hyper_params_choices = enumerate_parameters(hyper_params)
if random:
num_choices = len(self.hyper_params_choices)
assert random_max_run <= num_choices
sel = np.random.choice(range(num_choices),
size=(random_max_run, ),
replace=False)
self.hyper_params_choices = [
self.hyper_params_choices[s] for s in sel
]
self._define_masks()
def train_valid_search(self, X, y, v_X, v_y, param_grid, sample_weight=None, v_sample_weight=None, max_num_run=None, refit=False, verbose=1):
hyper_params = enumerate_parameters(param_grid)
hyper_params = np.random.choice(hyper_params, size=len(hyper_params), replace=False)
if max_num_run is None:
max_num_run = len(hyper_params)
best_params = None
best_score = 0
for r in range(max_num_run):
self.estimator = clone(self.estimator)
self.set_params(hyper_params[r])
self.fit(X, y, sample_weight=sample_weight)
if verbose>=2:
print hyper_params[r]
score = self.estimator.score(v_X, v_y, sample_weight=v_sample_weight)
if verbose>=2:
print score
if score > best_score:
best_score = score
best_params = hyper_params[r]
if verbose>=1:
print "best ------"
print best_params
print best_score
if sample_weight is None:
sample_weight = np.ones((X.shape[0], ))
if v_sample_weight is None:
v_sample_weight = np.ones((v_X.shape[0], ))
train_data = np.concatenate([X, y[:, None], sample_weight[:, None]], axis=1)
valid_data = np.concatenate([v_X, v_y[:, None], v_sample_weight[:, None]], axis=1)
all_data = np.concatenate([train_data, valid_data], axis=0)
np.random.shuffle(all_data)
self.set_params(best_params)
if refit:
self.estimator = clone(self.estimator)
self.fit(all_data[:, :-2], all_data[:, -2], sample_weight=all_data[:, -1])
def __init__(self,
hyper_params,
noise_generator,
inputs_dim,
prior_dim,
inputs_block,
attr_types,
random=False,
name="GDN"):
self.hyper_params = hyper_params
self.noise_generator = noise_generator
self.inputs_dim = inputs_dim
self.prior_dim = prior_dim
self.inputs_block = inputs_block
self.attr_types = attr_types
self.name = name
self.hyper_params_choices = enumerate_parameters(hyper_params)
if random:
num_choices = len(self.hyper_params_choices)
sel = np.random.choice(range(num_choices),
size=(num_choices, ),
replace=False)
self.hyper_params_choices = [
self.hyper_params_choices[s] for s in sel
]
self.set_model(self.hyper_params_choices[0])
def __init__(self, hyper_params, random=True, name=None):
self.name = name
self.hyper_params_choices = enumerate_parameters(hyper_params)
if random:
num_choices = len(self.hyper_params_choices)
sel = np.random.choice(range(num_choices),
size=(num_choices,), replace=False)
self.hyper_params_choices = [self.hyper_params_choices[s] for s in sel]
def __init__(self, method, hyper_params, random=False, random_max_run=1):
self.hyper_params_choices = enumerate_parameters(hyper_params)
if random:
num_choices = len(self.hyper_params_choices)
assert random_max_run <= num_choices
sel = np.random.choice(range(num_choices),
size=(random_max_run,), replace=False)
self.hyper_params_choices = [self.hyper_params_choices[s] for s in sel]
self.method = method
self.performance_records = []
def __init__(self, base_model, hyper_params, inputs_dim, random=True, name=None):
self.inputs_dim = inputs_dim
self.base_model = base_model
self.custom_objects = {}
if self.base_model == "Gaussian":
self.custom_objects["_mdn_gaussian_loss"] = self._mdn_gaussian_loss
elif self.base_model == "Poisson":
self.custom_objects["_mdn_poisson_loss"] = self._mdn_poisson_loss
elif self.base_model == "NegativeBinomial":
self.custom_objects["_mdn_negative_binomial_loss"] = self._mdn_negative_binomial_loss
self.hyper_params_choices = enumerate_parameters(hyper_params)
self.name = name
self.n_components = None
if random:
num_choices = len(self.hyper_params_choices)
sel = np.random.choice(range(num_choices),
size=(num_choices,), replace=False)
self.hyper_params_choices = [self.hyper_params_choices[s] for s in sel]