Back to DrivenData | Blog

Release of winning solutions?

Is it the norm for people to share their solutions/models once the competition has closed?
Really curious to see how people have approached it.

6 Likes

Just adding that I would like to see solutions as well!

1 Like

I’d also like hearing about other people’s solutions. There’s not much reason to do these contests if you don’t learn from other people.

Might as well share what I did…
Started with pre-trained ResNet and retrained it (allowing all the weights to be modified, not just the final layer as per normal transfer learning). Then used this as a feature extractor for various classifiers (tried XGBoost, SVM and a simple neural network). Ended up combining them into a simple averaging ensemble. Just about good enough to get in the top 50!

1 Like

Interesting. I’m afraid I didn’t have time or the hardware to fiddle with big neural networks so I just extracted some simple features from average colour, roof shape and location and ended up with a very similar score to you. Location was a pretty important feature. My solution sounds rather orthogonal to yours - perhaps some joint image neural network and location/shape features would do well.

1 Like