Keras Newsletter (June 2, 2023)

[haifeng]

Google I/O 2023 Recap

• Jonathan’s talk about KerasCV + KerasNLP on YouTube
  
  
• KerasCV+NLP featured in “What’s New In TensorFlow”
• On-device LLMs with Keras and Android on YouTube

KerasCV - 0.5 Release

• Unified Task API with KerasNLP
• YOLOV8 for object detection
• Vectorized augmentation layers
  • 10x speedup for many augmentation layers in graph mode
• Thanks to all our contributors!

KerasCV - Roadmap

• 0.5.1 coming in a few weeks
  • Primarily bug fixes and some docs cleanup
• Segmentation models
  • Segmentation mask support for more preprocessing layers
• 3D object detection models
  • Augmentation layers for LiDAR point clouds and 3D boxes

KerasNLP - 0.5 Release

• A good intro on keras.io - GPT2 Text Generation with KerasNLP
• Generative decoder models GPT2 and OPT.
  • Efficient, high-level text in → text out flow using XLA.
  • Low level samplers API for custom workflows.
• Masked language modeling tasks.
• Quality of life improvements.
  • Summary support including tokenizers and vocab counts.
  • Softmax output from classifiers.
• Thank you to our wide cast of contributors!!

KerasNLP - Roadmap

• New LLMs coming to KerasNLP.
  • Falcon, Pythia, MPT, etc.
• Seq2Seq models.
  • Bart, T5 and Whisper backbones on github today.
  • Full task workflows soon.
• Parameter efficient LORA training.
  • Example on keras.io incoming.
• Model parallel distributed training flows.