fastai snippets

Currently using fastai2.
Installation via pip install fastai2.

Oversampling

from fastai2.tabular.all import *
titanic = pd.read_csv('https://raw.githubusercontent.com/mwaskom/seaborn-data/master/titanic.csv')

cat_vars = ['survived', 'pclass', 'sex', 'embarked', 'class', 'who', 'adult_male', 'deck', 'embark_town', 'alone']
cont_vars = ['age', 'sibsp', 'parch', 'fare']
dep_var = 'alive'
procs = [Categorify, FillMissing, Normalize]
block_y = CategoryBlock()
splits = RandomSplitter()(range_of(titanic))

to = TabularPandas(titanic, procs=procs, cat_names=cat_vars, cont_names=cont_vars,
                   y_names=dep_var, block_y=block_y, splits=splits)

to.items

	survived	pclass	sex	age	sibsp	parch	fare	embarked	class	who	adult_male	deck	embark_town	alive	alone	age_na	sibsp_na	parch_na	fare_na
143	1	3	2	-0.799442	-0.464437	-0.467832	-0.509204	2	3	2	2	0	2	0	2	1	1	1	1
376	2	3	1	-0.568123	-0.464437	-0.467832	-0.498856	3	3	3	1	0	3	1	2	1	1	1	1
853	2	1	1	-1.030761	-0.464437	0.815107	0.166542	3	1	3	1	4	3	1	1	1	1	1	1
333	1	3	2	-1.030761	1.242489	-0.467832	-0.276367	3	3	2	2	0	3	0	1	1	1	1	1
791	1	2	2	-1.030761	-0.464437	-0.467832	-0.110793	3	2	2	2	0	3	0	2	1	1	1	1
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
206	1	3	2	0.202940	0.389026	-0.467832	-0.320865	3	3	2	2	0	3	0	1	1	1	1	1
654	1	3	1	-0.876548	-0.464437	-0.467832	-0.509204	2	3	3	1	0	2	0	2	1	1	1	1
802	2	1	2	-1.416293	0.389026	2.098045	1.834694	3	1	1	1	2	3	1	1	1	1	1	1
257	2	1	1	0.048727	-0.464437	-0.467832	1.141355	3	1	3	1	2	3	1	2	1	1	1	1
41	1	2	1	-0.182592	0.389026	-0.467832	-0.214277	3	2	3	1	0	3	0	1	1	1	1	1

891 rows × 19 columns

to.items.alive.value_counts().plot.bar();

dls = to.dataloaders()

learn = tabular_learner(dls, [200,100], metrics=accuracy)

The get_idxs method of DataLoaders samples the idxs to be used in an epoch of training:

learn.dls.get_idxs()[:10]

[412, 122, 391, 38, 310, 400, 316, 320, 57, 500]

len(learn.dls.get_idxs())

700

df = pd.DataFrame(1 / learn.dls.items.alive.value_counts())

weights = learn.dls.items.alive.apply(lambda x: df.to_dict()['alive'][x])

learn.dls.train.items.sample(n=700, weights=weights, replace=True).alive.value_counts()

0    352
1    348
Name: alive, dtype: int64

def sample_epoch(self):
    df = pd.DataFrame(1 / self.items.alive.value_counts())
    weights = self.items.alive.apply(lambda x: df.to_dict()['alive'][x])
    self.n = 700 # going for bigger epochs as this will require quite a bit of training
    oversampled_idxs = self.items.sample(n=self.n, weights=weights, replace=True).index
    return [np.where(self.items.index == i)[0][0] for i in oversampled_idxs]

learn.dls.train.get_idxs = types.MethodType(sample_epoch, learn.dls.train)

idxs = learn.dls.train.get_idxs()
learn.dls.train.items.iloc[idxs].alive.value_counts()

1    360
0    340
Name: alive, dtype: int64

learn.fit(10, 1e-3)

epoch	train_loss	valid_loss	accuracy	time
0	0.401528	0.673750	0.657303	00:00
1	0.231258	0.608156	0.792135	00:00
2	0.144609	0.453062	0.955056	00:00
3	0.098268	0.222973	1.000000	00:00
4	0.070314	0.066305	1.000000	00:00
5	0.052205	0.014175	1.000000	00:00
6	0.039717	0.002960	1.000000	00:00
7	0.030711	0.000851	1.000000	00:00
8	0.024077	0.000408	1.000000	00:00
9	0.019053	0.000268	1.000000	00:00

learn = tabular_learner(dls, [200,100], metrics=accuracy)

learn.fit(10, 1e-3)

epoch	train_loss	valid_loss	accuracy	time
0	0.328865	0.669253	0.775281	00:00
1	0.180982	0.614954	0.943820	00:00
2	0.110680	0.484824	0.977528	00:00
3	0.074929	0.271870	1.000000	00:00
4	0.053904	0.095419	1.000000	00:00
5	0.040015	0.020298	1.000000	00:00
6	0.030432	0.003644	1.000000	00:00
7	0.023582	0.000864	1.000000	00:00
8	0.018468	0.000273	1.000000	00:00
9	0.014615	0.000178	1.000000	00:00