V0.2.0

consists of PRs: #163 #161 #158 #152 #148 #142 #140 #138

• Refractor rafiki-web-admin with new UI
• Allow use of GPU for inference jobs
• Use colour images for IMAGE_CLASSIFICATION
• using kafka to transfer inference worker and predictor infomation
• Update model tuning for regression tasks and add new models
• Add support for tabular dataset and new models (Updated)
• add license to files
• Add support for tabular dataset and new models
• Add docs for speech recognition
• Add Speech Recognition model
• Add architecture tuning with ENAS & hyperparameter tuning with parameter sharing
• Change environment variable check method
• Add new models to Rafiki

V0.1.0

Overview of Changes

Consists of PRs: #119, #116, #115, #113, #108

• Rafiki’s data (i.e. model, job, user data) now persists across restarts
• Fixed “out-of-memory” error when large models are trained on Rafiki (by moving storage of model parameters from database to disk)
• App developers now can specify no. of GPUs to use for a train job, and during model training, each model is exclusively allocated a single GPU, if GPU is available
• Admins can now ban users
• Model developers can now delete models
• Stricter authentication checks on the back-end for management of models, users and jobs
• Better documentation (deployed at https://nginyc.github.io/rafiki/docs/0.1.0)
• User tokens now expire in 1h