Profile picture
Radek Osmulski @radekosmulski
, 8 tweets, 4 min read Read on Twitter
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
used a clever way of tuning predicted label probabilities based on correctly inferring the distribution of the test set. You can read the write up here: kaggle.com/c/quickdraw-do…. Congrats to the winners!

What are my personal takeaways from the competition? @kaggle is unbelievably
good for learning, esp if you join a competition early. At the same time, each competition (despite what it might seem initially) turns out to be a non trivial research project which takes a lot of time. In the words of @dhh from his new book, by saying yes to a single thing you
say no to a thousand other things. The opportunity costs can be high and need to be evaluated in the context of what one hopes to do down the road.

For over a year now I have compulsively jumped from a DL project to a DL project striving to maximize learning so maybe this is a
good time to start taking things more slowly and consider alternatives.

At the same time, @kaggle just launched a super exciting whale identification competition kaggle.com/c/humpback-wha… that I would love to get into🤷‍♂️Here is a great overview of a solution to an earlier version
of this competition: kaggle.com/martinpiotte/w…. A really great read and so many fun things one could do / try!
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to Radek Osmulski
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member and get exclusive features!

Premium member ($30.00/year)

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!