09_tabular kaggle competition: can't get my model learn, what I''m doing wrong?

Part 1 (2020)
1

Hi, I picked this MoA prediction competition as a way to practice what I’ve learned about tabular models on the fastai v2 course.

From my understanding I am correctly separating continuous from categorical variables, and creating a simple (300, 200) tabular model. I’ve separated the 10% randomly for validation, I’m using BCEWithLogitsLoss and I’m using the learning rate recommended by lr_find. I’ve also tried with weight decay.

However, the model seems to not learning much, and the validation loss barely changes from first epoch, with a score of ~0.69.

Please, could anyone shed some light on my errors?

I published this question as well at the competition discussion, as well as my notebook: https://www.kaggle.com/javiercarnero/moa-prediction-fastai

Thanks!

1 Like
2

Also interested; I had the same issue. The validation loss remained constant, but I went ahead and submitted the predictions from that model and got a reasonable (but by no means good) score. The model did train in spite of the learn.recoder.plot_loss() or the training table said, so I am questioning whether they are working properly for this tabular problem. (They probably are working; maybe I am just being a bozo.)

3

Discussion at kaggle hit it, the issue was with using y_range while using BCEWithLogits loss function, which apply sigmoid. Removing it from the learner did the trick.

1 Like