Poor Titanic Neural Net Results

Hello All!

I’m fairly new to the forums and fastai in general so please let me know if I can edit this post to make it more useful or more clear.


I’m working through the content for lesson 5 on my own here and I am able to get similar results to Jeremy using a linear regression, but when I try and implement a neural net, a deep network, or an ensemble, each time I get the same crappy accuracy which is about 20% worse than with the linear regression.

Even after re-watching the video and going over Jeremy’s notebook an (apparently) doing everything the same, I haven’t been able to get the same results. I must be doing something silly.

If anyone has the time to take a peak and provide some insight, it would be much appreciated.


Update: I’ve tried copying over all of the data processing code and all of the function declarations for the deep learning section and still I’m unable to get a result similar to to the one that was achieved here

Hi Logan
This sounds like a problem I ran into. I banged my head against the wall for a while.
You might be missing this step?

turn dependent variable into a column vector

trn_dep = trn_dep[:,None]
val_dep = val_dep[:,None]


Hey Chris,

Sorry for getting back to you late.

Thank you for the suggestion! Unfortunately, this did not fix my issue. :frowning:

Since I made this post I also tried going through Jeremy’s notebook and copying out every single data processing cell as well as all his function definitions and I still ended up with the same result. I even tried re-downloading the data to make sure I hadn’t modified it at some point.

When I moved on to using TabularPandas, I was able to start getting the same results again.

This is mildly frustrating but I’ve decided to move on since the point was not to make a perfect model from scratch, but to instead understand the process.