Tips on Improving Accuracy?

[lantean (Josh Coburn)]

I just finished lessons 1 & 2. In both lessons I’ve been trying to train a model to recognize 11 different species of fish. For my lesson 1 model, I hand selected 30 images of each species and got up to 88% accuracy on the validation set. After lesson 2, I downloaded around 350 images for each species and then cleaned using the techniques mentioned in the lecture & notebook. With all the additional data, I’m stuck at just better than 70% accuracy.

I’m not sure if this is normal, or if I’m doing something wrong. I guess 70% is still decent since random guessing would be under 10% accuracy, but it’s no where near the 90+% accuracy shown in the lecture. Should I keep working at this or should I keep moving with the lessons?

Notebooks can be found here:
https://github.com/JQCoburn/fish-id

[morgan (Morgan McGuire)]

Thats great progress given you’re at Lesson 2. More things you could try are below, some might be more advanced for where you are in your learning right now:

• more data
• increase regularisation with more augmentation, mixup is a great one to try, or playing around with weight decay
• Different optimizer, Ranger is my go-to at the moment
• Larger and/different architecture, EfficientNets are quite powerful and the smaller ones don’t take up too much memory

[lantean (Josh Coburn)]

Thank you @morgan! I’ll look into those! I appreciate the feedback!

[wcneill (Wesley Neill)]

Thanks, this was helpful for me too!