numbskulltypes.py

"""TODO."""

from __future__ import print_function, absolute_import
import numpy as np

# TODO (shared with DW): space optimization:
# 1. use smaller ints for some fields
# 2. replace a[x].length with a[x+1].offset - a[x].offset


Meta = np.dtype([('weights', np.int64),
                 ('variables', np.int64),
                 ('factors', np.int64),
                 ('edges', np.int64)])

Weight = np.dtype([("isFixed", np.bool),
                   ("parameterize",np.bool),       #标志位：权重是否需要参数化,参数化需要的参数个数后期在此处加
                   ("initialValue", np.float64),
                   ("a", np.float64),   #tau
                   ("b", np.float64)])  #alpha

Variable = np.dtype([("isEvidence", np.int8),
                     ("initialValue", np.int64),
                     ("dataType", np.int16),
                     ("cardinality", np.int64),
                     ("vtf_offset", np.int64)])
AlphaBound = np.dtype([("lowerBound", np.float64),  #用于Alpha参数控制
                    ("upperBound", np.float64)])
TauBound =  np.dtype([("lowerBound", np.float64),  #用于Alpha参数控制
                    ("upperBound", np.float64)])
SampleList = np.dtype([("vid", np.int64)])  #用于扩充0或1的证据以支持平衡化
Factor = np.dtype([("factorFunction", np.int16),
                   ("weightId", np.int64),
                   ("featureValue", np.float64),
                   ("arity", np.int64),
                   ("ftv_offset", np.int64)]) #此处的偏移量用来在fmap中找第n个因子的起始位置

FactorToVar = np.dtype([("vid", np.int64),
                        ("x", np.float64),
                        ("theta", np.float64),   #新加的，以传入对应的theta值
                        ("dense_equal_to", np.int64)])

VarToFactor = np.dtype([("value", np.int64),
                        ("factor_index_offset", np.int64),
                        ("factor_index_length", np.int64)])

UnaryFactorOpt = np.dtype([('vid', np.int64),
                           ('weightId', np.int64)])