Whale, Dolphin, Shark classifier has low accuracy rate

Part 1 (2019)
1

Hey guys!

I’ve been trying to improve my accuracy rate with my classification algorithm but it’s really low ~89% at the best right now. Additionally, the error rate jumps which I thought was related to overfitting so I ran fewer epochs before unfreezing and that didn’t improve accuracy. Then I tried reducing the learning rate and that helped slightly but then the error was dropping too slowly.

Here’s the data I’m using: https://www.dropbox.com/s/nks6hkxb7alemul/dolphin_whale_shark.tar.gz?dl=0

Here’s my notebook: https://gist.github.com/dasickis/f000e398597807984774b7acba76ee30

2

Looks like you might want to do some data cleaning. For example, there’s no way it’s going to know that an airplane painted like a whale is a whale if you don’t have other training examples like that. Similar story for drawings of dolphins.

You don’t have that many data points so each one it misclassifies affects your accuracy quite a bit.

All told, just judging by what I can see from your notebook, it looks like you’re doing pretty well.

2 Likes
3

Sounds good. Will spend time today clearing the training data.

4

That worked! We went up to 96% accuracy rate which makes total sense. Thank you!