How do start with Planet: Understanding the Amazon from Space on Kaggle considering I have limited knowledge on satellite images and have completed Part 1.
There’s really nothing to know - just download the images and get started using the techniques you learnt! Just use the jpegs, which are standard 3 channel images.
The only difference to what you’ve done already is that these images have multiple labels. So you’ll need to remove any softmax layer, and think a bit about how best to represent the data.
I have been trying to implement various classifiers in pytorch and keras. When I see any implementations online and find the fast ai implementation not only the accuracy is better but there is a huge difference even with the same architecture. So I tried to implement planet amazon from space in pytorch. I tried to replicate most of the things apart from data augmentation and cyclical learning I could from the fast ai version with fbeta score 93.03, The results were very bad. Can anyone help me out with what I am doing wrong.
Links to kaggle kernel: