NVIDIA released its latest version of Clara Train and Deploy application frameworks. From AutoML to workflow manager for priority scheduling these releases bring state-of-the-art capabilities to AI development and deployment in medical imaging.
Clara Train v3.0
New AI-Assisted Annotation Features:
- Interactive annotation kick-started with a novel semi-automated deep learning-based approach using foreground and background clicks to accelerate AI-ready datasets even when a pre-trained model is not available
- Customizable annotation workflow with the ability to bring your own models or chain of models to an annotation workflow
New Training Framework Features:
- An AutoML module to search for optimal parameters when tuning and testing a new model, increasing efficiency and maximizing GPU utilization
- Clara Federated Learning updates with bring your own privacy policies and aggregators
- enabling research collaboration in building robust AI models while preserving privacy
- New Keras based imaging pipeline with faster loader functions and Novagrad optimizer
Additional Resources:
GTC Digital Clara Developer Day Tutorial for AutoML
GTC Digital Clara Developer Day Tutorial for Federated Learning
Clara Deploy v0.5
New Platform Features:
- Supported pipeline composition with strongly typed operator interface enabling pre-runtime validation of pipelines, compatibility of concatenated operators in terms of data type, and allocation of memory for the pipeline using FAST I/O via CPDriver.
- Built-in scheduler to help manage and allocate resources to execute pipeline jobs.
- An AI model management tool with the ability to store and manage models locally through user inputs, pull models in from external stores via direct download services to create and manage model catalogs.
- A CLI Load Generator to simulate expected hospital workloads to ensure both the architecting of hardware and software can support the development of application pipelines.
- A Fast I/O feature to provide an interface to memory resources that are accessible by all operators running in the same pipeline. Memory resources can be used for efficient, zero-copy sharing and passing of data between operators.
Additional Resources:
GTC Digital Clara Developer Day tutorial Scalable and Modular Deployment with Clara Deploy
To learn more about Scalable and Modular AI Deployment with Clara Deploy Application Framework – download the whitepaper.
New Reference Application Pipelines:
- Prostate Segmentation
- Multi-AI Organ Segmentation
- 3D Cropping using shared memory
- DeepStream Video Batch
Related blogs:
Building a Medical Imaging AI Ecosystem Using Clara SDKs
Clara Developer Day: Clara Train SDK Perform Walkthrough and Deep Dive
