2/ In 2016 Sarada founded the Perth ML Group to help others learn.
How can the community support you? 🤗
It can...
✅ help you set up your environments 🧑💻
✅ provide technically-sound answers to challenging questions 💡
✅ make learning more fun! 🥳
3/ What are some tips for community participation?
✅ explaining things to others will help you learn 🦉
✅ it's okay to be anxious about sharing your answers publicly - DMs are always an option 📨
✅ experiment with various approaches and learn in a way that suits you best 💡
4/ This is purely an appreciation tweet 🙂
Sarada is part of the @fastdotai community. But I have not been aware of the many communities she is contributing to and the many ways in which she is making the world a better place.
Honored to learn from her and be inspired 🙏
• • •
Missing some Tweet in this thread? You can try to
force a refresh
This is how little code it takes to implement a siamese net using @fastdotai and @pytorch.
I share this because I continue to be amazed.
Here is a refactored version that will be easier to change
The models above were my naive adaptations of the simple siamese network concept from cs.utoronto.ca/~gkoch/files/m… (screenshot on the left below) to a CNN setting.
On the right is the network from the Learning Deep Convolutional Feature Hierarchies section but using pretrained models
Now that the Quickdraw! competition is over, I added my solution to the repository github.com/radekosmulski/…. It is a model consisting of outputs of res50, incv4 and an rnn along with country code embeddings fed to an fc classifier. The most interesting aspect of the architecture
is how little code wiring everything together required thx to the @fastdotai lib. I replaced the classifier during training (while retaining the other weights) which the @fastdotai / @pytorch combo made very easy to accomplish.
In some sense there is nothing particularly
interesting about my solution above and beyond the basic motions of building a model. For various reasons I didn't get to the point of applying competition specific insights. The model is also partially trained
The winning solution was a CNN ensemble done using LightGBM. It also