There is no .process() for a DataLoader generated from your âoriginalâ dl used in training (as in a dl created from calling dls.test_dl). You can do that for your training dl though. Even after calling .process() and then creating a new .test_dl using my test data, the dataset is still unencoded (which makes sense, but I just wanted to mention that).
Thank you, Iâll have a look! However, I am a bit confused. I just created a TabularDataLoaders with a list of two variables as y_names, and y_block=RegressionBlock(), and things seem to be working with MSE as loss function.
I did not even need to adjuts the n_out argument of RegressionBlock, which I was thinking at the beginning that could be used to define the number of output activations that you wanted in the regression.
One extra question: I am not sure if I remember correctly, but: did fastai1 provide a way to automatically log transform the dependent variables? Is that provided as a ready-to-use Procs in fastai2?
EDIT: Sorry I just saw that Jeremy does it manually in one of the fastai2 lessons
In the meanwhile I am using the show method from TabularPandas, but instead of using the whole validation dataset, I manually subset it with the indices given by Interpretation.top_losses:
@vrodriguezf the issue comes with _pre_show_batch. Itâs not returning the y's or the outs as it should be. Iâll file a bug report (As @sgugger may not be able to get to this for a bit). In the meantime nice workaround!
You should be able to use it for regression IIRC (it was a collaboration project and I believe we tackled it). Otherwise, yes permutation importance would be good to use too always I canât recall if I had it set up specifically for classification or not, but it should be straightforward to adjust for regression too (just adjust what metric/loss function itâs using to generate itâs differences and possibly change how it uses them. IE MSE should favor a smaller number vs hinder such as accuracy)