Help with Tabular learning

Part 1 (2020)
1

Hi

I’m using the Tabular learner for classification. I’m at a loss why the accuracy is not improving here.

image
image332×590 45.7 KB

Any suggestions? Thank you!!!
Here’s why I picked the learning rate

image

2

It’s strange. I seems that your model isn’t overfitting although reach only needs 1 epoch to achieve the top accuracy. Also, it’s strange that accuracy never changes.

My thoughts are:

  1. Your model may be learning to predict a subset of classes as accuracy doesn’t change. For example, the most common class. You can check by showing network predictions & top losses. It’s very strange. Likely, you are doing something wrong. Review your code.
  2. Your data has some order and you aren’t shuffling it before creating the train / validation split. For example, first N rows contains class A, N+M class B, etc. Try to use RandomShuffle for splitting.
  3. Your model is underfitting (unlikely as accuracy doesn’t change). Try a bigger model.
  4. You have a messy data / very complicated problem (unlikely as accuracy doesn’t change). Try a bigger model.

I hope that it helps.

3

Hi everyone, I am trying to post a new topic but I literally have no idea how. I have looked through every part of this website for a ‘post new topic’ type of button but can’t find one. I can reply which is why I am asking you

4

Found a problem with my script.
In the line below I didn’t add a “y_block=CategoryBlock” before. Now I think the nnn knows I want to do classification and not regression.

to = TabularPandas(data, procs=procs, cat_names = cat_vars, cont_names = cont_vars, y_names='myfavy', splits=splits, y_block=CategoryBlock)

The accuracy looks much better, but it’s still not improving with training :frowning:
image

1 Like
5

Well it is improving, just very slowly :smiley: The losses are decreasing nicely. I would try running it for more epochs and/or with a higher learning rate.

6

Welcome! :slight_smile: I think you will have to read some posts and visit a few times in the next days to unlock the “new topic” button.
In the meantime, there are separate topics for every lesson of the course, maybe your question will fit in there!