Predicting for Test data on TabularDataLoader learner

I have created TabularDataLoader using:

dls = TabularDataLoaders.from_csv(‘abc.csv’, y_names=“Y”, y_block = CategoryBlock,
cat_names = [‘A’, ‘B’,
‘C’,‘D’,‘E’, ‘E’],
cont_names = [‘G’, ‘H’],
procs = [Categorify, FillMissing, Normalize], split=RandomSplitter(seed=42))

I then use tabular learner to train with the above dls. Post that I use:

testdl = learn.dls.test_dl(testdf, process=true) # where testdf is the test data
preds, _ = learn.get_preds(dl=testdl)

I get errors on missing values. I thought same procs that were used in the TabularDataLoaders for training would be used. Is that not true? It’s not just missing values, I wonder if none of the procs are applied on test data and so the actual predictions are wrong. I tried separately filling the missing values and the error disappeared, but the predictions appear to be incorrect. The accuracy post training is 85%, but when I try to predict for a different set of data, the accuracy is very low <50%. I’ve never been able to solve this problem of test data for predictions using exactly the same procs as used in training dls. Is there a standard way to get this right or is this a gap?