Twitter | Search | |
Jason Antic
Software guy, currently digging deep into GANs to do some cool photo colorization and restoration! Author of DeOldify.
1,016
Tweets
2,215
Following
1,560
Followers
Tweets
Jason Antic retweeted
Deirdre Apr 15
since turning 30 i spend a lot of time wistfully looking out windows like an old fisherman haunted by memories of the sea
Reply Retweet Like
Jason Antic retweeted
Jeremy Howard Apr 14
The heart of a training loop with callbacks. By aligning the training code and callback code, you can see exactly what's going on in each. Formatting code for understanding is too important to leave to automated tools or hard and fast rules.
Reply Retweet Like
Jason Antic retweeted
Jeremy Howard Apr 11
"Pretty nice" is quite the understatement. This is a wonderful in-depth discussion of the weird interactions between batchnorm, weight decay, and learning rate, including a fascinating experiment that shows that you can entirely replace weight decay with learning rate changes,
Reply Retweet Like
Jason Antic retweeted
helena sarin Apr 5
the lesson i'm deep learning is that whatever looks nice and charming done by a human hand would look mechanical and forced when is done with GAN however similar the texture is... the quest is for a new kind of mark making - unabashedly done with machine but unique and beautiful
Reply Retweet Like
Jason Antic Apr 5
omg this is brilliant
Reply Retweet Like
Jason Antic retweeted
Sam Finlayson Apr 4
Holy heckfire the replies to this thread are terrifying. It is completely unconscionable to have patients being treated by people in these physical/mental states.
Reply Retweet Like
Jason Antic Apr 3
Jeremy really has a knack for giving cool talks. Here’s another example of that!
Reply Retweet Like
Jason Antic retweeted
Sylvain Gugger Apr 1
Thanks to you, default tokenization in is now 10 times faster!
Reply Retweet Like
Jason Antic retweeted
Sebastian Ruder Apr 2
Pretrained language models are not only applicable to natural language but also to other domains where sequences have an underlying structure, such as genomics. We can get better performance with more meaningful token representations (e.g. using k-mers instead of nucleotides).
Reply Retweet Like
Jason Antic retweeted
Sam Stiyer Apr 1
Reply Retweet Like
Jason Antic retweeted
helena sarin Mar 30
Reply Retweet Like
Jason Antic retweeted
Jeremy Howard Mar 28
Pretrained ULMFiT language models for 10 Indian languages!
Reply Retweet Like
Jason Antic retweeted
Jeremy Howard Mar 27
Keras has been an inspiration to us at and to so many others! Thank you! :D
Reply Retweet Like
Jason Antic retweeted
Jeremy Howard Mar 23
This is a fun application of the superres method from lesson 7 - turning line drawings into shaded pictures!
Reply Retweet Like
Jason Antic retweeted
Jeremy Howard Mar 24
. : "If you were starting today, what is the one resource you'd use to get started in AI?" : "" 😊😁😎 (The other 66 questions and answers are interesting too...)
Reply Retweet Like
Jason Antic retweeted
Emma Wedekind 🐞 Mar 22
I don’t believe a web developer should be tested on low-level computer science knowledge. Are certain aspects of CS important to development? Yes. Should web devs be grilled on how to convert decimal to binary? No. Can we please stop tricking devs in interviews?
Reply Retweet Like
Jason Antic retweeted
Dan Abramov Mar 23
I’ll start: most component unit tests create more friction than what they’re worth.
Reply Retweet Like
Jason Antic retweeted
Danielle Perez Mar 22
Remember when speaking in a fake voice and lying about inventing technology just meant you were in middle school
Reply Retweet Like
Jason Antic retweeted
Gerry McBride Mar 22
Parents, never leave kids unsupervised when they’re watching YouTube. You can never tell what they’re going to click through to. I left my son unattended with my phone and somehow he ended up watching the WASSUP Bud ad from 20 years ago and now he will not stop fucking saying it
Reply Retweet Like
Jason Antic retweeted
Sylvain Gugger Mar 21
Here is the recording of the talk about the callback system in fastai I gave yesterday hosted by for the meetup in NYC. My part begins at 7:32.
Reply Retweet Like