Misunderstanding size of final layer in tabular_leaner

fastai fastai dev
1

I’m playing around with the fast.ai library V1 and I don’t understand why the last layer doesn’t include the out_sz but the last emb_szs1. I’m running the tabular example in /example/tabular.

tabular
tabular.png1134×832 81.5 KB

I think it should be 2 but not 100 right ? It is a binary classification model

When I see in 009_rossmann notebook, a regression model, the output size is 1.

rossman
rossman.png918×765 68.3 KB

Thank you in advance.

2

It’s a bug that has been fixed. You’ll see it in the next release of the fastai library on conda/pip (or maybe already if you update).

1 Like
3

Thank you @sgugger. I’m currently at fast.ai 1.0.5 which is the lastest released version. Do you mean if I reinstall fast.ai version for developper the bug will be fixed ?

4

[Update] I uninstalled fast.ai conda version and reinstalled the developer version following the instructions on github. It works now :smiley: Thank you so much.

1 Like