def partial_fit_base(self, X, y):
check_is_fitted(self, "base_model_")
batch_indices = generate_batch(
X, self.autograd_config.get("batch_size", 32))
esp = 1e-11 # where should this live?
step_size = self.autograd_config.get("step_size", 0.05)
callback = (None if self.autograd_config.get("verbose", False) else
simple_callback)
num_iters = self.autograd_config.get("num_iters", 1000)
nclass = self.n_classes_
model_dump = self.base_model_.booster_.dump_model()
trees_ = [m["tree_structure"] for m in model_dump["tree_info"]]
trees_params = multi_tree_to_param(X, y, trees_)
model_ = gbm_gen(trees_params[0], X, trees_params[2], trees_params[1],
False, 2)
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
training_gradient_fun = grad(training_loss)
param_ = adam(
training_gradient_fun,
trees_params[0],
callback=callback,
step_size=step_size,
num_iters=num_iters,
)
self.base_param_ = copy.deepcopy(trees_params)
self.partial_param_ = param_
self.is_partial = True
return self
def partial_fit_param(self, X, y):
check_is_fitted(self, "base_model_")
check_is_fitted(self, "base_param_")
check_is_fitted(self, "partial_param_")
batch_indices = generate_batch(
X, self.autograd_config.get("batch_size", 32))
esp = 1e-11 # where should this live?
step_size = self.autograd_config.get("step_size", 0.05)
callback = (None if self.autograd_config.get("verbose", False) else
simple_callback)
num_iters = self.autograd_config.get("num_iters", 1000)
nclass = self.n_classes_
if nclass == 2:
y_ohe = y
else:
y_ohe = LabelBinarizer().fit_transform(y)
if nclass > 2:
model_ = gbm_gen(
self.base_param_[0],
X,
self.base_param_[2],
self.base_param_[1],
True,
nclass,
)
def training_loss(weights, idx=0):
# Training loss is the negative log-likelihood of the training labels.
t_idx_ = batch_indices(idx)
preds = model_(weights, X[t_idx_, :])
loglik = -np.sum(np.log(preds + esp) * y_ohe[t_idx_, :])
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
else:
model_ = gbm_gen(
self.base_param_[0],
X,
self.base_param_[2],
self.base_param_[1],
False,
2,
)
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
training_gradient_fun = grad(training_loss)
param_ = adam(
training_gradient_fun,
self.partial_param_,
callback=callback,
step_size=step_size,
num_iters=num_iters,
)
self.partial_param_ = param_
self.is_partial = True
return self