FastAI QuickStart Tabular Data

Source

The Beginning

Imports

# python
from functools import partial

# fastai
from fastai.tabular.all import (
    Categorify,
    FillMissing,
    Normalize,
    TabularDataLoaders,
    URLs,
    accuracy,
    tabular_learner,
    untar_data,
)

# pypy
from tabulate import tabulate

import numpy
import pandas

# my stuff
from graeae import Timer

table = partial(tabulate, tablefmt="orgtbl", headers="keys")
TIMER = Timer()

The Middle

The Data

We're using the Adult Data Set, which has an unfortunate title but is a dataset built from 1994 census data to predict whether a person has an income greater than $50,000 a year.

path = untar_data(URLs.ADULT_SAMPLE)
DATA_PATH = path/"adult.csv"
data = pandas.read_csv(DATA_PATH)

numerical = data.select_dtypes(include=[numpy.number])
non_numerical = data.select_dtypes(exclude=[numpy.number])

print(table(numerical.describe()))
  age fnlwgt education-num capital-gain capital-loss hours-per-week
count 32561 32561 32074 32561 32561 32561
mean 38.5816 189778 10.0798 1077.65 87.3038 40.4375
std 13.6404 105550 2.573 7385.29 402.96 12.3474
min 17 12285 1 0 0 1
25% 28 117827 9 0 0 40
50% 37 178356 10 0 0 40
75% 48 237051 12 0 0 45
max 90 1.4847e+06 16 99999 4356 99 
print(table(non_numerical.describe()))
  workclass education marital-status occupation relationship race sex native-country salary
count 32561 32561 32561 32049 32561 32561 32561 32561 32561
unique 9 16 7 15 6 5 2 42 2
top Private HS-grad Married-civ-spouse Prof-specialty Husband White Male United-States <50k
freq 22696 10501 14976 4073 13193 27816 21790 29170 24720

The column names don't really make clear what some things are, but since this is a quickstart I'll ignore their meaning but note that it was useful to split the data up by numeric and non-numeric types becaus when you build the TabularDataLoader you should specify the numeric and categorical column names. The fastai example only specifies some of the columns but I'll dump them all in and see what happens.

numeric_columns = numerical.columns.to_list()
categorical_columns = non_numerical.columns.to_list()[:-1]

The Data Loader

The original quickstart uses the TabularDataLoaders class to load batches of data for training, along with some pre-processing classes to encode the categorical data to make it numeric, fill in the missing values, and normalize the values so their ranges will match.

TARGET = "salary"

loader = TabularDataLoaders.from_csv(
    DATA_PATH, path=path, y_names=TARGET,
    cat_names = categorical_columns,
    cont_names = numeric_columns,
    procs = [Categorify, FillMissing, Normalize])

The Learner

learner = tabular_learner(loader, metrics=accuracy)

with learner.no_bar() as nobu, TIMER as tim:
    learner.fit_one_cycle(2)

Started: 2022-11-06 17:09:17.344678
[0, 0.37480291724205017, 0.35229262709617615, 0.8412162065505981, '00:02']
[1, 0.3569386303424835, 0.34605613350868225, 0.8421375751495361, '00:02']
Ended: 2022-11-06 17:09:23.994030
Elapsed: 0:00:06.649352

The Learned

Since the last column salary is the target we'll have to drop it before training the model on the data.

unsalaried = data.drop(["salary"], axis=1)

test_set = learner.dls.test_dl(unsalaried)

row, classifications, probabilities = learner.predict(
    data.iloc[0])

Sources