StateFarm - Get decent accuracy/validation accuracy, terrible kaggle score

Part 1 (2017)
1

I’ve gone ahead and created a simple sequential network with no CNNs in it to test out before I did a CNN. After some effort I have it generating pretty decent results even with a good-enough validation accuracy.

I completed the submission with just the linear model to see how my score would fare - and it resulted in one of the worst possible loss scores on kaggle.

The file can be viewed here.

What am I doing wrong?

Additional problem:
Some of you may note that I use Shuffle=True. When I set this to Shuffle=False as per suggestion of some posts in this forum, my accuracy always says “0.1000” no matter how many epochs I do, or what my learning rate is. Which makes me wonder if I broke something or I’m missing something entirely.

Thanks for your time!

2

You can ignore this topic. I managed to confuse myself thoroughly by trying the StateFarm dataset PRIOR to watching lesson 3, and have learned a ton of lessons I need to digest and try out before bothering anyone with this crappy linear model.

I also now know what I was doing wrong - for people in the future, I did not separate out the drivers. someone on this forum provided a great script to handle this for you. You essentially train the network to detect drivers.