TabModel

This is an implementation created by Ignacio Oguiza (oguiza@timeseriesAI.co) based on fastai’s TabularModel.

We built it so that it’s easy to change the head of the model, something that is particularly interesting when building hybrid models.

source

TabHead


def TabHead(
    emb_szs, n_cont, c_out, layers:NoneType=None, fc_dropout:NoneType=None, y_range:NoneType=None, use_bn:bool=True,
    bn_final:bool=False, lin_first:bool=False, act:ReLU=ReLU(inplace=True), skip:bool=False
):

Basic head for tabular data.

source

TabBackbone


def TabBackbone(
    emb_szs, n_cont, embed_p:float=0.0, bn_cont:bool=True
):

Same as nn.Module, but no need for subclasses to call super().__init__

source

TabModel


def TabModel(
    emb_szs, n_cont, c_out, layers:NoneType=None, fc_dropout:NoneType=None, embed_p:float=0.0, y_range:NoneType=None,
    use_bn:bool=True, bn_final:bool=False, bn_cont:bool=True, lin_first:bool=False, act:ReLU=ReLU(inplace=True),
    skip:bool=False
):

Basic model for tabular data.

from fastai.tabular.core import *
from tsai.data.tabular import *

path = untar_data(URLs.ADULT_SAMPLE)
df = pd.read_csv(path/'adult.csv')
# df['salary'] = np.random.rand(len(df)) # uncomment to simulate a cont dependent variable
procs = [Categorify, FillMissing, Normalize]
cat_names = ['workclass', 'education', 'marital-status', 'occupation', 'relationship', 'race']
cont_names = ['age', 'fnlwgt', 'education-num']
y_names = ['salary']
y_block = RegressionBlock() if isinstance(df['salary'].values[0], float) else CategoryBlock()
splits = RandomSplitter()(range_of(df))
pd.options.mode.chained_assignment=None
to = TabularPandas(df, procs=procs, cat_names=cat_names, cont_names=cont_names, y_names=y_names, y_block=y_block, splits=splits, inplace=True, 
                   reduce_memory=False)
to.show(5)
tab_dls = to.dataloaders(bs=16, val_bs=32)
b = first(tab_dls.train)
test_eq((b[0].shape, b[1].shape, b[2].shape), (torch.Size([16, 7]), torch.Size([16, 3]), torch.Size([16, 1])))

	workclass	education	marital-status	occupation	relationship	race	education-num_na	age	fnlwgt	education-num	salary
20505	Private	HS-grad	Married-civ-spouse	Sales	Husband	White	False	47.0	197836.0	9.0	<50k
28679	Private	HS-grad	Married-civ-spouse	Craft-repair	Husband	White	False	28.0	65078.0	9.0	>=50k
11669	Private	HS-grad	Never-married	Adm-clerical	Not-in-family	White	False	38.0	202683.0	9.0	<50k
29079	Self-emp-not-inc	Bachelors	Married-civ-spouse	Prof-specialty	Husband	White	False	41.0	168098.0	13.0	<50k
7061	Private	HS-grad	Married-civ-spouse	Adm-clerical	Husband	White	False	31.0	243442.0	9.0	<50k

tab_model = build_tabular_model(TabModel, dls=tab_dls)
b = first(tab_dls.train)
test_eq(tab_model.to(b[0].device)(*b[:-1]).shape, (tab_dls.bs, tab_dls.c))
learn = Learner(tab_dls, tab_model, splitter=ts_splitter)
p1 = count_parameters(learn.model)
learn.freeze()
p2 = count_parameters(learn.model)
learn.unfreeze()
p3 = count_parameters(learn.model)
assert p1 == p3
assert p1 > p2 > 0