The documentation of Orange package didn't cover all the details. Table._init__(Domain, numpy.ndarray)
works only for int
and float
according to lib_kernel.cpp
.
They really should provide an C-level interface for pandas.DataFrames
, or at least numpy.dtype("str")
support.
Update: Adding table2df
, df2table
performance improved greatly by utilizing numpy for int and float.
Keep this piece of script in your orange python script collections, now you are equipped with pandas in your orange environment.
Usage: a_pandas_dataframe = table2df( a_orange_table )
, a_orange_table = df2table( a_pandas_dataframe )
Note: This script works only in Python 2.x, refer to @DustinTang 's answer for Python 3.x compatible script.
import pandas as pd
import numpy as np
import Orange
#### For those who are familiar with pandas
#### Correspondence:
#### value <-> Orange.data.Value
#### NaN <-> ["?", "~", "."] # Don't know, Don't care, Other
#### dtype <-> Orange.feature.Descriptor
#### category, int <-> Orange.feature.Discrete # category: > pandas 0.15
#### int, float <-> Orange.feature.Continuous # Continuous = core.FloatVariable
#### # refer to feature/__init__.py
#### str <-> Orange.feature.String
#### object <-> Orange.feature.Python
#### DataFrame.dtypes <-> Orange.data.Domain
#### DataFrame.DataFrame <-> Orange.data.Table = Orange.orange.ExampleTable
#### # You will need this if you are reading sources
def series2descriptor(d, discrete=False):
if d.dtype is np.dtype("float"):
return Orange.feature.Continuous(str(d.name))
elif d.dtype is np.dtype("int"):
return Orange.feature.Continuous(str(d.name), number_of_decimals=0)
else:
t = d.unique()
if discrete or len(t) < len(d) / 2:
t.sort()
return Orange.feature.Discrete(str(d.name), values=list(t.astype("str")))
else:
return Orange.feature.String(str(d.name))
def df2domain(df):
featurelist = [series2descriptor(df.icol(col)) for col in xrange(len(df.columns))]
return Orange.data.Domain(featurelist)
def df2table(df):
# It seems they are using native python object/lists internally for Orange.data types (?)
# And I didn't find a constructor suitable for pandas.DataFrame since it may carry
# multiple dtypes
# --> the best approximate is Orange.data.Table.__init__(domain, numpy.ndarray),
# --> but the dtype of numpy array can only be "int" and "float"
# --> * refer to src/orange/lib_kernel.cpp 3059:
# --> * if (((*vi)->varType != TValue::INTVAR) && ((*vi)->varType != TValue::FLOATVAR))
# --> Documents never mentioned >_<
# So we use numpy constructor for those int/float columns, python list constructor for other
tdomain = df2domain(df)
ttables = [series2table(df.icol(i), tdomain[i]) for i in xrange(len(df.columns))]
return Orange.data.Table(ttables)
# For performance concerns, here are my results
# dtndarray = np.random.rand(100000, 100)
# dtlist = list(dtndarray)
# tdomain = Orange.data.Domain([Orange.feature.Continuous("var" + str(i)) for i in xrange(100)])
# tinsts = [Orange.data.Instance(tdomain, list(dtlist[i]) )for i in xrange(len(dtlist))]
# t = Orange.data.Table(tdomain, tinsts)
#
# timeit list(dtndarray) # 45.6ms
# timeit [Orange.data.Instance(tdomain, list(dtlist[i])) for i in xrange(len(dtlist))] # 3.28s
# timeit Orange.data.Table(tdomain, tinsts) # 280ms
# timeit Orange.data.Table(tdomain, dtndarray) # 380ms
#
# As illustrated above, utilizing constructor with ndarray can greatly improve performance
# So one may conceive better converter based on these results
def series2table(series, variable):
if series.dtype is np.dtype("int") or series.dtype is np.dtype("float"):
# Use numpy
# Table._init__(Domain, numpy.ndarray)
return Orange.data.Table(Orange.data.Domain(variable), series.values[:, np.newaxis])
else:
# Build instance list
# Table.__init__(Domain, list_of_instances)
tdomain = Orange.data.Domain(variable)
tinsts = [Orange.data.Instance(tdomain, [i]) for i in series]
return Orange.data.Table(tdomain, tinsts)
# 5x performance
def column2df(col):
if type(col.domain[0]) is Orange.feature.Continuous:
return (col.domain[0].name, pd.Series(col.to_numpy()[0].flatten()))
else:
tmp = pd.Series(np.array(list(col)).flatten()) # type(tmp) -> np.array( dtype=list (Orange.data.Value) )
tmp = tmp.apply(lambda x: str(x[0]))
return (col.domain[0].name, tmp)
def table2df(tab):
# Orange.data.Table().to_numpy() cannot handle strings
# So we must build the array column by column,
# When it comes to strings, python list is used
series = [column2df(tab.select(i)) for i in xrange(len(tab.domain))]
series_name = [i[0] for i in series] # To keep the order of variables unchanged
series_data = dict(series)
print series_data
return pd.DataFrame(series_data, columns=series_name)