Tabular data preprocessing¶

In [ ]:

from fastai.gen_doc.nbdoc import *
from fastai.tabular import *
from fastai import *

Overview¶

This package contains the basic class to define a transformation for preprocessing dataframes of tabular data, as well as basic TabularTransform. Preprocessing includes things like

replacing non-numerical variables by categories, then their ids,
filling missing values,
normalizing continuous variables.

In all those steps we have to be careful to use the correspondance we decide on our training set (which id we give to each category, what is the value we put for missing data, or how the mean/std we use to normalize) on our validation or test set. To deal with this, we use a speciall class called TabularTransform.

The data used in this document page is a subset of the adult dataset. It gives a certain amount of data on individuals to train a model to predict wether their salary is greater than $50k or not.

In [ ]:

path = untar_data(URLs.ADULT_SAMPLE)
df = pd.read_csv(path/'adult.csv')
train_df, valid_df = df.iloc[:800].copy(), df.iloc[800:1000].copy()
train_df.head()

Out[ ]:

	age	workclass	fnlwgt	education	education-num	marital-status	occupation	relationship	race	sex	capital-gain	capital-loss	hours-per-week	native-country	>=50k
0	49	Private	101320	Assoc-acdm	12.0	Married-civ-spouse	NaN	Wife	White	Female	0	1902	40	United-States	1
1	44	Private	236746	Masters	14.0	Divorced	Exec-managerial	Not-in-family	White	Male	10520	0	45	United-States	1
2	38	Private	96185	HS-grad	NaN	Divorced	NaN	Unmarried	Black	Female	0	0	32	United-States	0
3	38	Self-emp-inc	112847	Prof-school	15.0	Married-civ-spouse	Prof-specialty	Husband	Asian-Pac-Islander	Male	0	0	40	United-States	1
4	42	Self-emp-not-inc	82297	7th-8th	NaN	Married-civ-spouse	Other-service	Wife	Black	Female	0	0	50	United-States	0

We see it contains numerical variables (like age or education-num) as well as categorical ones (like workclass or relationship). The original dataset is clean, but we removed a few values to give examples of dealing with missing variables.

In [ ]:

cat_names = ['workclass', 'education', 'marital-status', 'occupation', 'relationship', 'race', 'sex', 'native-country']
cont_names = ['age', 'fnlwgt', 'education-num', 'capital-gain', 'capital-loss', 'hours-per-week']

Transforms for tabular data¶

In [ ]:

show_doc(TabularProc, doc_string=False)

`class` `TabularProc`[source]

TabularProc(cat_names:StrList, cont_names:StrList)

Base class for creating transforms for dataframes with categorical variables cat_names and continuous variables cont_names. Note that any column not in one of those lists won't be touched.

In [ ]:

show_doc(TabularProc.__call__)

`call`[source]

call(df:DataFrame, test:bool=False)

Apply the correct function to df depending on test.

This simply calls apply_test if test or apply_train otherwise. Those functions apply the changes in place.

In [ ]:

show_doc(TabularProc.apply_train, doc_string=False)

`apply_train`[source]

apply_train(df:DataFrame)

Must be implemented by an inherited class with the desired transformation logic.

In [ ]:

show_doc(TabularProc.apply_test, doc_string=False)

`apply_test`[source]

apply_test(df:DataFrame)

If not implemented by an inherited class, defaults to calling apply_train.

The following TabularTransform are implemented in the fastai library. Note that the replacement from categories to codes as well as the normalization of continuous variables are automatically done in a TabularDataset.

In [ ]:

show_doc(Categorify, doc_string=False)

`class` `Categorify`[source]

Categorify(cat_names:StrList, cont_names:StrList) :: TabularProc

Changes the categorical variables in cat_names in categories. Variables in cont_names aren't affected.

In [ ]:

show_doc(Categorify.apply_train, doc_string=False)

`apply_train`[source]

apply_train(df:DataFrame)

Transforms the variable in the cat_names columns in categories. The category codes are the unique values in these columns.

In [ ]:

show_doc(Categorify.apply_test, doc_string=False)

`apply_test`[source]

apply_test(df:DataFrame)

Transforms the variable in the cat_names columns in categories. The category codes are the ones used for the training set, new categories are replaced by NaN.

In [ ]:

tfm = Categorify(cat_names, cont_names)
tfm(train_df)
tfm(valid_df, test=True)

Since we haven't changed the categories by their codes, nothing visible has changed in the dataframe yet, but we can check that the variables are now categorical and view their corresponding codes.

In [ ]:

train_df['workclass'].cat.categories

Out[ ]:

Index([' ?', ' Federal-gov', ' Local-gov', ' Private', ' Self-emp-inc',
       ' Self-emp-not-inc', ' State-gov', ' Without-pay'],
      dtype='object')

The test set will be given the same category codes as the training set.

In [ ]:

valid_df['workclass'].cat.categories

Out[ ]:

Index([' ?', ' Federal-gov', ' Local-gov', ' Private', ' Self-emp-inc',
       ' Self-emp-not-inc', ' State-gov', ' Without-pay'],
      dtype='object')

In [ ]:

show_doc(FillMissing, doc_string=False)

`class` `FillMissing`[source]

FillMissing(cat_names:StrList, cont_names:StrList, fill_strategy:FillStrategy=<FillStrategy.MEDIAN: 1>, add_col:bool=True, fill_val:float=0.0) :: TabularProc

Transform that fills the missing values in cont_names. cat_names variables are left untouched (their missing value will be raplced by code 0 in the TabularDataset). fill_strategy is adopted to replace those nans and if add_col is True, whenever a column c has missing values, a column named c_nan is added and flags the line where the value was missing.

In [ ]:

show_doc(FillMissing.apply_train, doc_string=False)

`apply_train`[source]

apply_train(df:DataFrame)

Fills the missing values in the cont_names columns.

In [ ]:

show_doc(FillMissing.apply_test, doc_string=False)

`apply_test`[source]

apply_test(df:DataFrame)

Fills the missing values in the cont_names columns with the ones picked during train.

In [ ]:

train_df[cont_names].head()

Out[ ]:

	age	fnlwgt	education-num	capital-gain	capital-loss	hours-per-week
0	49	101320	12.0	0	1902	40
1	44	236746	14.0	10520	0	45
2	38	96185	NaN	0	0	32
3	38	112847	15.0	0	0	40
4	42	82297	NaN	0	0	50

In [ ]:

tfm = FillMissing(cat_names, cont_names)
tfm(train_df)
tfm(valid_df, test=True)
train_df[cont_names].head()

Out[ ]:

	age	fnlwgt	education-num	capital-gain	capital-loss	hours-per-week
0	49	101320	12.0	0	1902	40
1	44	236746	14.0	10520	0	45
2	38	96185	10.0	0	0	32
3	38	112847	15.0	0	0	40
4	42	82297	10.0	0	0	50

Values issing in the education-num column are replaced by 10, which is the median of the column in train_df. Categorical variables are not changed, since nan is simply used as another category.

In [ ]:

valid_df[cont_names].head()

Out[ ]:

	age	fnlwgt	education-num	capital-gain	hours-per-week
800	45	96975	10.0	0	40
801	46	192779	10.0	15024	60
802	36	376455	10.0	0	38
803	25	50053	10.0	0	45
804	37	164526	10.0	0	40

In [ ]:

%reload_ext autoreload
%autoreload 2
%matplotlib inline

In [ ]:

show_doc(FillStrategy, alt_doc_string='Enum flag represents determines how `FillMissing` should handle missing/nan values', arg_comments={
    'MEDIAN':'nans are replaced by the median value of the column',
    'COMMON': 'nans are replaced by the most common value of the column',
    'CONSTANT': 'nans are replaced by `fill_val`'
})

`FillStrategy`

Enum = [MEDIAN, COMMON, CONSTANT]

Enum flag represents determines how FillMissing should handle missing/nan values

MEDIAN: nans are replaced by the median value of the column
COMMON: nans are replaced by the most common value of the column
CONSTANT: nans are replaced by fill_val

In [ ]:

show_doc(Normalize, doc_string=False)

`class` `Normalize`[source]

Normalize(cat_names:StrList, cont_names:StrList) :: TabularProc

In [ ]:

show_doc(Normalize.apply_train, doc_string=False)

`apply_train`[source]

apply_train(df:DataFrame)

Computes the means and stds on the continuous variables of df then normalizes those columns.

In [ ]:

show_doc(Normalize.apply_test, doc_string=False)

`apply_test`[source]

apply_test(df:DataFrame)

Use the means and stds stored to normalize the continuous columns of df.

Tabular data preprocessing¶

Overview¶

Transforms for tabular data¶

`class` `TabularProc`[source]

`call`[source]

`apply_train`[source]

`apply_test`[source]

`class` `Categorify`[source]

`apply_train`[source]

`apply_test`[source]

`class` `FillMissing`[source]

`apply_train`[source]

`apply_test`[source]

`FillStrategy`

`class` `Normalize`[source]

`apply_train`[source]

`apply_test`[source]

Undocumented Methods - Methods moved below this line will intentionally be hidden¶

New Methods - Please document or move to the undocumented section¶

Tabular data preprocessing¶

Overview¶

Transforms for tabular data¶

class TabularProc[source]

__call__[source]

apply_train[source]

apply_test[source]

class Categorify[source]

apply_train[source]

apply_test[source]

class FillMissing[source]

apply_train[source]

apply_test[source]

`FillStrategy`

class Normalize[source]

apply_train[source]

apply_test[source]

Undocumented Methods - Methods moved below this line will intentionally be hidden¶

New Methods - Please document or move to the undocumented section¶

`class` `TabularProc`[source]

`call`[source]

`apply_train`[source]

`apply_test`[source]

`class` `Categorify`[source]

`apply_train`[source]

`apply_test`[source]

`class` `FillMissing`[source]

`apply_train`[source]

`apply_test`[source]

`class` `Normalize`[source]

`apply_train`[source]

`apply_test`[source]