Tabular batch prediction with fastai v2

[ajka]

I have trained a tabular_learner on a tabular dataset, and I would now like to invoke the model with a test dataset.

I attempted to do this using learner.get_preds:

But I ran into a couple of errors:

---

I then decided to try just predicting row by row, using pandas.DataFrame.apply:

But I got a KeyError:

KeyError: "None of [Index([ <column names printed here> ],\n      dtype='object', name=<index column name printed here>)] are in the [columns]"

---

Then, I decided to just try iterating over the rows in the dataframe:

But this is extremely slow and it seems like it would take most of a day to complete executing.

---

What is the correct way to get batch predictions for a test dataset with a TabularLearner in fastai v2? If anyone can point me in the right direction, I’d really appreciate it!

[stefan-ai (Stefan Josef)]

Hi @ajka

This should work (I used it for text_classifier_learner but it should be the same for tabular):

test_dl = learner.dls.test_dl(test_binary_features)
predictions = learner.get_preds(dl=test_dl)

[ajka]

That worked perfectly. Thanks again, @stefan-ai!

[henry090 (Turgut)]

Thanks for your reply. However, I have a question. For the new dataset when we do not know the output (label column) of the data, then the predictions for the label part throws None. I think something is wrong here. Because with standard predict method we get the probability+label (for example “yes”/“no”). What do you think?

[[1]]
   workclass  education  marital-status  ...    fnlwgt  education-num  salary
0        5.0        8.0             3.0  ... -0.837419       0.753672     1.0

[1 rows x 11 columns]

[[2]]
tensor(1)

[[3]]
tensor([0.4342, 0.5658])

Update: probabilities with predict and your suggestion are right but can we get the labels as well?

[stefan-ai (Stefan Josef)]

@henry090: learn.predict() and learn.get_preds() give different outputs. learn.predict() returns predicted class names, predicted class index and predicted probabilities. learn.get_preds() on the other hand returns predicted probabilities and true classes. So for a new dataset without labels, it’s correct that the second output returns None.

You can then take the predicted probabilities, get the predicted class index using argmax and the corresponding class name by indexing into dls.vocab (or dls.vocab[1] in NLP).

[muellerzr (Zachary Mueller)]

@henry090

but can we get the labels as well

fastinference will do this for you BTW: See here

(for some reason it’s not showing it but you should be able to pass with_decoded=True to get_preds)

[henry090 (Turgut)]

@stefan-ai @muellerzr Both answers are very helpful! dls.vocab for column names and with_decoded=True which will append labels as well! Thanks very much!

[henry090 (Turgut)]

Hi @stefan-ai , @muellerzr . When we specify as metric roc auc binary, then while predicting a single row I get :

  ValueError: Only one class present in y_true. ROC AUC score is not defined in that case. 

Is there a way to skip AUC score evaluation while predicting single rows?

[muellerzr (Zachary Mueller)]

@henry090 you should be able to remove it from learn.metrics