-
Notifications
You must be signed in to change notification settings - Fork 1k
特征准入
songyue1104 edited this page Aug 12, 2019
·
2 revisions
-
- 概率准入
参数:feature_add_probability
取值:0到1的浮点数
说明:表示该feature准入的概率,只对新特征有效,如果一个特征已经准入则不走该逻辑
eg:
embs = xdl.embedding(name = 'probility_var',
sparse_input = inputs,
initializer = xdl.Zeros(),
emb_dim = 8,
feature_dim = 1,
vtype='hash',
feature_add_probability = 0.5)
-
- 基于CBF的精确准入
参数:cbf
取值:大于0的整数
说明:表示特征准入的阈值,只有特征出现次数大于阈值才能参与训练,否则返回全零的embedding
eg:
embs = xdl.embedding(name = 'probility_var',
sparse_input = inputs,
initializer = xdl.Zeros(),
emb_dim = 8,
feature_dim = 1,
vtype='hash',
cbf = 2)