Hey, all!
Today I’ve got an implementation of WideResNet for you. Trying to figure out what these papers are saying and how to implement them in fast.ai is really teaching me a lot. Some takeaways:
- Darknet really is fast – really
- Though I did not test this myself, if I extrapolate from WRN’s performance in this notebook, I can see how it would perform better against fresh ResNets with a similar number of parameters; based on this and other results from the paper, I will seriously consider widening a network before deepening if I ever feel the need to increase the number of layers beyond 50
- As you can see, train_loss consistently hovers above valid_loss in later epochs, while valid_loss continues to drop; my suspicion is that the dropout layers are really helping the network generalize – bears further investigation
Enjoy!