Weighted data loading of tabular data

Link to forum discussion. [1] https://forums.fast.ai/t/weighted-dataloaders-with-tabularpandas/102057/2 [2] https://forums.fast.ai/t/using-weighteddl-with-tabulardata/80789

I don't find anyone successfully using tabular data with sample weights.

Is your feature request related to a problem? Please describe. It's necessary to handle class imbalance in tabular data.

Describe the solution you'd like Most straightforward solution would be adding wgts to TabularDataLoaders: https://docs.fast.ai/tabular.data.html

I tried to use TabularPandas in combination with calling dataloaders(dl_type=WeightedDL, bs=64, wgts=...), but it leads to an error described in [2]. I tried to fix it myself but it's too complex for me. I observed two issues:

[] operator ends up with used on the tabular data which causes looking up the columns than the rows.
collate that happens later supports numpy, tensor, and so on only.

Describe alternatives you've considered Instead of using WeightedDL, I can try sampling my data frame before calling dataloader. But that's going to be very cumbersome.

Additional context

Steps to reproduce:

from fastai.tabular.all import *

path = untar_data(URLs.ADULT_SAMPLE)
df = pd.read_csv(path/'adult.csv', skipinitialspace=True)
df.head()
cat_names = ['workclass', 'education', 'marital-status', 'occupation', 'relationship']
cont_names = ['age', 'fnlwgt', 'education-num']
procs = [Categorify, FillMissing, Normalize]

to = TabularPandas(
    df, procs=procs,
    cat_names=cat_names, cont_names=cont_names, y_names='race', 
    y_block = CategoryBlock())

# works ok
dls = to.dataloaders(bs=20)
dls.show_batch()

# fails.
# "Could not do one pass in your dataloader, there is something wrong in it. Please see the stack trace below"
dls = to.dataloaders(bs=20, dl_type=WeightedDL, wgts=[1.0] * len(df))

fastai / fastai

Weighted data loading of tabular data #3860