关于 Pandas 多索引 HDFStore 的磁盘索引

On disk indexing of Pandas multiindexed HDFStore

为了提高性能并减少内存占用,我正在尝试读取在 Pandas 中创建的多索引 HDFStore。原来的商店很大,但问题可以用一个类似但更小的例子重现。

df = pd.DataFrame([0.25, 0.5, 0.75, 1.0],
                      index=['Item0', 'Item1', 'Item2', 'Item3'], columns=['Values'])

df = pd.concat((df.iloc[:],df.iloc[:]), axis=0,names=['Item','N'], 
               keys = ['Items0','Items1'])

df.to_hdf('hdfs.h5', 'df', format='table',mode='w',complevel= 9,complib='blosc',data_columns=True) 

store = pd.HDFStore('hdfs.h5', mode= 'r')

store.select('df',where='Item="Items0"')

这应该是 return 子索引的值,但是它 return 是一个错误

> ValueError: The passed where expression: Item="Items0"
>             contains an invalid variable reference
>             all of the variable refrences must be a reference to
>             an axis (e.g. 'index' or 'columns'), or a data_column
>             The currently defined references are: index,iron,columns

指数是:

store['df'].index

> MultiIndex(levels=[['Items0', 'Items1'], ['Item0', 'Item1', 'Item2',
> 'Item3']],
>            labels=[[0, 0, 0, 0, 1, 1, 1, 1], [0, 1, 2, 3, 0, 1, 2, 3]],
>            names=['Item', 'N'])

谁能解释一下可能是什么原因?或者应该如何正确完成...

对我来说,如果删除 data_columns=True:

df.to_hdf('hdfs3.h5', 'df', format='table',mode='w',complevel= 9,complib='blosc') 
store = pd.HDFStore('hdfs3.h5', mode= 'r')
print (store.select('df','Item="Items0"'))
              Values
Item   N            
Items0 Item0    0.25
       Item1    0.50
       Item2    0.75
       Item3    1.00

尝试将 data_columns=True 替换为 data_columns=df.columns.tolist()

演示:

原始 MultiIndex DF:

In [2]: df
Out[2]:
              Values
Item   N
Items0 Item0    0.25
       Item1    0.50
       Item2    0.75
       Item3    1.00
Items1 Item0    0.25
       Item1    0.50
       Item2    0.75
       Item3    1.00

使用 data_columns=df.columns.tolist():

将其保存到 HDF5
In [3]: df.to_hdf('c:/temp/hdfs.h5','df',format='t',mode='w',complevel=9,complib='blosc',data_columns=df.columns.tolist())

In [4]: df.columns.tolist()
Out[4]: ['Values']

从 HDF 商店中选择:

In [5]: store = pd.HDFStore('c:/temp/hdfs.h5')

索引级别和 Values 列现在都已编入索引,可以在 where=<query> 参数中使用:

In [6]: store.select('df',where='Item="Items0" and Values in [0.5, 1]')
Out[6]:
              Values
Item   N
Items0 Item1     0.5
       Item3     1.0

In [7]: store.select('df',where='N="Item3" and Values in [0.5, 1]')
Out[7]:
              Values
Item   N
Items0 Item3     1.0
Items1 Item3     1.0

店家信息:

In [8]: store.get_storer('df').table
Out[8]:
/df/table (Table(8,), shuffle, blosc(9)) ''
  description := {
  "index": Int64Col(shape=(), dflt=0, pos=0),
  "N": StringCol(itemsize=5, shape=(), dflt=b'', pos=1),
  "Item": StringCol(itemsize=6, shape=(), dflt=b'', pos=2),
  "Values": Float64Col(shape=(), dflt=0.0, pos=3)}
  byteorder := 'little'
  chunkshape := (2427,)
  autoindex := True
  colindexes := {
    "Values": Index(6, medium, shuffle, zlib(1)).is_csi=False,
    "index": Index(6, medium, shuffle, zlib(1)).is_csi=False,
    "Item": Index(6, medium, shuffle, zlib(1)).is_csi=False,
    "N": Index(6, medium, shuffle, zlib(1)).is_csi=False}

存储索引级别:

In [9]: store.get_storer('df').levels
Out[9]: ['Item', 'N']

注意: 如果您只是省略 data_columns 参数,那么只有索引会在 HDF 存储中建立索引,所有其他列将不可搜索:

演示:

In [19]: df.to_hdf('c:/temp/NO_data_columns.h5', 'df', format='t',mode='w',complevel=9,complib='blosc')

In [20]: store = pd.HDFStore('c:/temp/NO_data_columns.h5')

In [21]: store.select('df',where='N == "Item3"')
Out[21]:
              Values
Item   N
Items0 Item3     1.0
Items1 Item3     1.0

In [22]: store.select('df',where='N == "Item3" and Values == 1')
---------------------------------------------------------------------------
...
skipped
...

ValueError: The passed where expression: N == "Item3" and Values == 1
            contains an invalid variable reference
            all of the variable refrences must be a reference to
            an axis (e.g. 'index' or 'columns'), or a data_column
            The currently defined references are: N,index,Item,columns

更新:

What is the real difference in putting data_columns=df.columns.tolist() ?

In [18]: fn = r'd:/temp/a.h5'

In [19]: df.to_hdf(fn,'dc_true',data_columns=True,format='t',mode='w',complevel=9,complib='blosc')

In [20]: df.to_hdf(fn,'dc_cols',data_columns=df.columns.tolist(),format='t',complevel=9,complib='blosc')

In [21]: store = pd.HDFStore(fn)

In [22]: store
Out[22]:
<class 'pandas.io.pytables.HDFStore'>
File path: d:/temp/a.h5
/dc_cols            frame_table  (typ->appendable_multi,nrows->8,ncols->3,indexers->[index],dc->[N,Item,Values])
/dc_true            frame_table  (typ->appendable_multi,nrows->8,ncols->3,indexers->[index],dc->[Values])

In [23]: store.get_storer('dc_true').table.colindexes
Out[23]:
{
    "Values": Index(6, medium, shuffle, zlib(1)).is_csi=False,
    "index": Index(6, medium, shuffle, zlib(1)).is_csi=False}

In [24]: store.get_storer('dc_cols').table.colindexes
Out[24]:
{
    "Item": Index(6, medium, shuffle, zlib(1)).is_csi=False,  # <- missing when `data_columns=True`
    "N": Index(6, medium, shuffle, zlib(1)).is_csi=False,     # <- missing when `data_columns=True`
    "Values": Index(6, medium, shuffle, zlib(1)).is_csi=False,
    "index": Index(6, medium, shuffle, zlib(1)).is_csi=False}

所以区别在于索引列的索引方式