def _split_next(self):
self._bridge.start(self._bridge.new_iter_id())
split_info = tree_pb2.SplitInfo()
self._bridge.receive_proto(
self._bridge.current_iter_id, 'split_info') \
.Unpack(split_info)
node = self._nodes[split_info.node_id]
node.left_child = self._add_node(node.node_id)
left_child = self._nodes[node.left_child]
left_child.weight = float('nan')
node.right_child = self._add_node(node.node_id)
right_child = self._nodes[node.right_child]
right_child.weight = float('nan')
self._num_leaves += 1
if split_info.feature_id >= 0:
self._set_node_partition(node, split_info)
self._bridge.send_proto(
self._bridge.current_iter_id, 'follower_split_info',
tree_pb2.SplitInfo(left_samples=left_child.sample_ids,
right_samples=right_child.sample_ids))
else:
node.is_owner = False
left_child.sample_ids = list(split_info.left_samples)
right_child.sample_ids = list(split_info.right_samples)
self._bridge.commit()
return left_child, right_child, split_info
def _find_split_and_push(self, node):
split_info = tree_pb2.SplitInfo(
node_id=node.node_id, gain=-1)
for fid, (grad_hist, hess_hist) in \
enumerate(zip(node.grad_hists, node.hess_hists)):
sum_g = sum(grad_hist)
sum_h = sum(hess_hist)
left_g = 0.0
left_h = 0.0
nan_g = grad_hist[-1]
nan_h = hess_hist[-1]
for i in range(len(grad_hist) - 2):
left_g += grad_hist[i]
left_h += hess_hist[i]
self._compare_split(
split_info, True, fid, i,
left_g + nan_g, left_h + nan_h,
sum_g - left_g - nan_g, sum_h - left_h - nan_h)
self._compare_split(
split_info, False, fid, i,
left_g, left_h,
sum_g - left_g, sum_h - left_h)
self._split_candidates.put((-split_info.gain, split_info))
return split_info.gain, split_info
def _split_next(self):
self._bridge.start(self._bridge.new_iter_id())
_, split_info = self._split_candidates.get()
node = self._nodes[split_info.node_id]
node.left_child = self._add_node(node.node_id)
left_child = self._nodes[node.left_child]
left_child.weight = split_info.left_weight
node.right_child = self._add_node(node.node_id)
right_child = self._nodes[node.right_child]
right_child.weight = split_info.right_weight
self._num_leaves += 1
if split_info.feature_id < self._binned.features.shape[1]:
self._set_node_partition(node, split_info)
self._bridge.send_proto(
self._bridge.current_iter_id, 'split_info',
tree_pb2.SplitInfo(node_id=split_info.node_id,
feature_id=-1,
left_samples=left_child.sample_ids,
right_samples=right_child.sample_ids))
else:
node.is_owner = False
fid = split_info.feature_id - self._binned.features.shape[1]
self._bridge.send_proto(
self._bridge.current_iter_id, 'split_info',
tree_pb2.SplitInfo(node_id=split_info.node_id,
feature_id=fid,
split_point=split_info.split_point,
default_left=split_info.default_left))
split_info.feature_id = -1
follower_split_info = tree_pb2.SplitInfo()
self._bridge.receive_proto(
self._bridge.current_iter_id, 'follower_split_info') \
.Unpack(follower_split_info)
left_child.sample_ids = list(follower_split_info.left_samples)
right_child.sample_ids = list(follower_split_info.right_samples)
self._bridge.commit()
return left_child, right_child, split_info
def _find_split_and_push(self, node):
max_gain = -1
max_fid = None
split_point = None
left_weight = None
right_weight = None
lam = self._l2_regularization
for fid, (grad_hist, hess_hist) in \
enumerate(zip(node.grad_hists, node.hess_hists)):
sum_g = sum(grad_hist[:-1])
sum_h = sum(hess_hist[:-1])
left_g = 0.0
left_h = 0.0
for i in range(len(grad_hist[:-1]) - 1):
left_g += grad_hist[i]
left_h += hess_hist[i]
right_g = sum_g - left_g
right_h = sum_h - left_h
gain = left_g*left_g/(left_h + lam) + \
right_g*right_g/(right_h + lam) - \
sum_g*sum_g/(sum_h + lam)
if gain > max_gain:
max_gain = gain
max_fid = fid
split_point = i
left_weight = -left_g / (left_h + lam)
right_weight = -right_g / (right_h + lam)
split_info = tree_pb2.SplitInfo(
node_id=node.node_id,
gain=max_gain,
feature_id=max_fid,
split_point=split_point,
left_weight=left_weight * self._learning_rate,
right_weight=right_weight * self._learning_rate)
self._split_candidates.put((-max_gain, split_info))
return max_gain, split_info
