Two Years in the Life of a Deep Learning Library in Kotlin
summary
A retrospective of building KotlinDL, a Deep Learning library for Kotlin that layers Kotlin APIs on top of TensorFlow’s compute core to enable training and fine-tuning neural networks on the JVM with TensorFlow/PyTorch‑level performance. Covers the two-year engineering journey from a one-class MVP to a multi-module open-source project with hundreds of classes and contributors, including topics like interfacing with native and Python ecosystems (JNI vs JavaCPP), parsing model weights, running MLOps out of the box, coordinating unpaid contributors, and using company resources to grow an OSS project.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 2:09Speaker background
- 9:48Early deep learning experiments
- 15:47Prototyping the library
- 22:49First release feedback
- 27:43Expanding functionality
- 31:33Industry tool integration
- 36:28New neural architectures
- 40:20Community contributions
- 44:28Pivot to ONNX runtime
- 48:37Final release
- 50:21Future outlook
related talks
49mDeep Dive into Deep Learning with KotlinDL
44mKotlin strives for Deep Learning
47mDeep Learning: a practical introduction from an Android dev perspective
45mKickstarting Kotlin Culture: The Journey from Java to Kotlin
45mMeanwhile in Kotlin: what happened over the last year
06mOpening Address
33mIntegrating Kotlin into a Poliglot JVM Environment
01h 04mWhat you always wanted to know about Deep Learning, but were afraid to ask