Tabular: set up new train and valid `dls` with initially prepared `to` object

fastai
1

HI,

I try to find the way to set up new train and valid dls with initially prepared to object.

Below are the two versions of the code, first is a simple example from tutorial and it works well,
second is an attempt to get dataloaders with initially prepared to.

I expected both snippets to work exactly the same, but there is zero learning improvement learning with the second snippet.

This basic example works as expected:

df is a pd.DataFrame

splits = RandomSplitter(valid_pct=0.2)(range_of(df))
cont_names = df.select_dtypes(['float64','int64']).columns.tolist()
cat_names =  df.select_dtypes(['bool','category']).columns.tolist()
cat_names = [c for c in cat_names if c not in ['sold']]
splits = RandomSplitter(valid_pct=0.2)(range_of(df))
to = TabularPandas(df, procs=[Categorify, FillMissing,Normalize],
                   cat_names = cat_names,
                   cont_names = cont_names,
                   y_names='sold',
                   splits=splits,
                   inplace=True,
                   reduce_memory=False)

dls = to.dataloaders(bs=64*50).cuda()
learn = tabular_learner(dls, metrics=accuracy)
learn.fit_one_cycle(40, lr_max=1e-04)

This code causing causing accuracy metric to bounce around 0.5 all the time :

# create `to` object with setup data

cont_names = df.select_dtypes(['float64','int64']).columns.tolist()
cat_names =  df.select_dtypes(['bool','category']).columns.tolist()
cat_names = [c for c in cat_names if c not in ['sold']]

 splits = RandomSplitter(valid_pct=0.2)(range_of(df))
 to = TabularPandas(df, procs=[Categorify, FillMissing,Normalize],
               cat_names = cat_names,
               cont_names = cont_names,
               y_names='sold',
               splits=splits,
               inplace=True,
               reduce_memory=False)



splits = RandomSplitter(valid_pct=0.2)(range_of(df))

train_ds = df[df.index.isin(splits[0])]
valid_ds = df[dff.index.isin(splits[1])]

new_train_to = to.new(train_ds)
new_valid_to = to.new(valid_ds)
new_train_to.process()
new_valid_to.process()

trn_dl = TabDataLoader(new_train_to,bs=64*50)
val_dl = TabDataLoader(new_valid_to,bs=64*50)

dls =  DataLoaders(trn_dl, val_dl).cuda()
learn = tabular_learner(dls, metrics=accuracy)
learn.fit_one_cycle(40, lr_max=1e-04)

Thank you!

1 Like
2

Hi Dmitry,

Have you managed to solve the problem? I have similar.

Thanks

3

Hi krasin, the reason was that you need to shuffle train dataloader. Dataloaders constructer does that for you automatically, but ones you do is yourself with TabDataLoader, you have to add shuffle = True

new_train_to = to.new(train_ds)
new_valid_to = to.new(valid_ds)
new_train_to.process()
new_valid_to.process()

trn_dl  =TabDataLoader(new_train_to,bs=64*50, shuffle=True)
val_dl = TabDataLoader(new_valid_to,bs=64*50)

dls =  DataLoaders(trn_dl, val_dl).cuda()