Hello everyone! KTensorFlow 2.0 is out :tada: — a ...

# feed

[Sergey Kurochkin]

Hello everyone!
KTensorFlow 2.0 is out 🎉 — a Kotlin Multiplatform library for running TensorFlow Lite (LiteRT) models on Android and iOS from shared code.

Some of you may remember 1.x. Version 2.0 grew from a thin wrapper into a toolkit for on-device ML, all in common Kotlin:

New in 2.0
• Vision module — a cross-platform Image (Bitmap on Android, CGImage on iOS) with resizing, cropping, rotation and letterboxing done by native platform APIs, plus conversion to tensors, bounding boxes and NMS
• Coroutines module — runSuspend, processFlow and processFlowDropping, which skips camera frames while the model is busy, so live detection doesn't fall behind
• Zero-copy tensor views — reshape, transpose, slice and others share memory instead of copying it
• Model metadata — signature names, named outputs and resizeInput
• Thread safety — interpreters and delegates are AutoCloseable, safe to call from several threads and safe to close while running

Still there from 1.x
• Loading .tflite models from the shared module, with Compose Resources and Moko Resources support
• GPU and NPU acceleration
• Tensors with arithmetic, casting, argmax / argmin
• Declarative pipelines for preprocessing and postprocessing

The repo includes a sample that runs live object detection from the camera on both platforms with the same shared code.

Next on the roadmap
• Converting camera frames into existing tensors, so live inference allocates no memory per frame
• Passing context such as letterbox info through pipelines
• More built-in pipeline stages and tensor operations
• Text processing utilities

👉 github.com/kursor1337/KTensorFlow

What started as my graduation project has become a real passion. I'd love to hear your feedback and ideas, and contributions are welcome — and a ⭐ helps a lot!

1005.mov