https://kotlinlang.org logo
Title
z

zaleslaw

04/27/2021, 10:33 AM
Today I want to share with you my presentation about an upcoming 0.2 KotlinDL release (0.2-alpha-1 artifact is available now on Maven Central). Currently, the KotlinDL framework can boast that this is the only way to construct and train complex neural networks on JVM, such as VGG, ResNet, or MobileNet. There is also support for transfer learning for the popular models trained in Keras (or available in Keras. applications). For image preprocessing, several functions are available that allow you to avoid complex and routine work on the JVM. Code from the demo: https://github.com/zaleslaw/KotlinDL-Demo/tree/demo-1 Library: https://github.com/JetBrains/KotlinDL

https://www.youtube.com/watch?v=jCFZc97_XQU

P.S. Support our project with a star on Github!
👍 9
🔥 1
р

Ролан

04/27/2021, 2:08 PM
@zaleslaw why is it not possible to build those architectures with dl4j?
I am sorry if I missed something.
z

zaleslaw

04/27/2021, 2:09 PM
Never see Mobilenet built with dl4j
р

Ролан

04/27/2021, 2:11 PM
what does mobilenet exatly use that you cannot reproduce?
The reason I am asking actually is because I would like to compile a set of examples accross several frameworks on JVM to test for perfomance and compare on usability
z

zaleslaw

04/27/2021, 2:18 PM
So, I have a production experience with Dl4j in 2017-18, as I remember it was really hard to describe there huge models, prepocess data, load weights from keras for transfer learning. Now, I don't know anyone who use it in production and have no information about project roadmap, and I did not keep in mind this framework when writing the marketing words above, but means modern frameworks as Java API for tf, PyTorch Java API from Facebook, tribuo and Java bindings for onnx runtime
But thanks for the link with examples, I will revisit it to keep in mind the current level of this framework
р

Ролан

04/27/2021, 2:21 PM
OK fair enough, the java API for tf and torch are for inference only I guess (at least the "documented" part indeed)
Was Samuel Audet already working on dl4j in 2018?
z

zaleslaw

04/27/2021, 2:24 PM
It's possible, he is old player in this field, and highly likely worked in Skymind
It will be very interesting to measure performance in some standard use cases, I tried to do the same a year ago and discovered a lot of undocumented things, will be interesting for me to collaborate/discuss it (list of the frameworks, methodology, possible cases)
р

Ролан

04/27/2021, 2:29 PM
@zaleslaw thank you a lot - I plan (with hopefully a student) to start working on that very seriously in July.
z

zaleslaw

04/27/2021, 2:30 PM
So, ping me, if you need a collaborator here
р

Ролан

04/27/2021, 2:30 PM
Wonderful - I will certainly do 🤗