def predict(self, X):
check_is_fitted(self, 'base_model_')
if not self.is_partial:
return self.base_model_.predict(X)
else:
multi_class = self.n_classes_ > 2
model_ = gbm_gen(self.partial_param_, X, self.base_param_[2],
self.base_param_[1], multi_class, self.n_classes_)
preds = model_(self.partial_param_, X)
if self.n_classes_ > 2:
return np.argmax(preds, axis=1)
else:
return np.round(sigmoid(preds))
def predict_proba(self, X):
check_is_fitted(self, 'base_model_')
if not self.is_partial:
return self.base_model_.predict_proba(X)
else:
multi_class = self.n_classes_ > 2
model_ = gbm_gen(self.partial_param_, X, self.base_param_[2],
self.base_param_[1], multi_class, self.n_classes_)
preds = model_(self.partial_param_, X)
if self.n_classes_ > 2:
return preds
else:
pred_positive = sigmoid(preds)
return np.stack([1 - pred_positive, pred_positive], axis=-1)
def training_loss(weights, idx=0):
# Training loss is the negative log-likelihood of the training labels.
t_idx_ = batch_indices(idx)
preds = sigmoid(model_(weights, X[t_idx_, :]))
label_probabilities = preds * y[t_idx_] + (1 - preds) * (1 -
y[t_idx_])
# print(label_probabilities)
loglik = -np.sum(np.log(label_probabilities))
num_unpack = 3
reg = 0
# reg_l1 = np.sum(np.abs(flattened)) * 1.
for idx_ in range(0, len(weights), num_unpack):
param_temp_ = weights[idx_:idx_ + num_unpack]
flattened, _ = weights_flatten(param_temp_[:2])
reg_l1 = np.sum(np.abs(flattened)) * 1.0
reg += reg_l1
return loglik + reg