February 2021 Summaries

Clarifai's Release 7.2 introduces advanced deep training templates and an enhanced polygon labeling interface in Scribe, aimed at improving model building and image annotation processes. The release includes the Clarifai_ResNet_AngularMargin, a model designed for high-performance face recognition and visual search tasks by using limited sample images to distinguish between numerous people or objects, and the Clarifai_EfficientDet, an efficient object detector built for scalable, fast, and lightweight performance on edge devices. Clarifai_EfficientDet utilizes a compound scaling method to enhance model efficiency by uniformly scaling various network components. Additionally, the update features improved polygon labeling tools that facilitate the efficient insertion and editing of multi-node polygon images for more precise object annotation in labeling projects.

Clarifai chose to build its API on gRPC due to its cutting-edge performance, enabling lightweight microservices that efficiently manage thousands of machine learning model instances and deliver scalable MLOps. The use of gRPC, which leverages the Protobuf message format and is designed for HTTP/2, ensures low latency and high throughput communications, significantly reducing network constraints compared to JSON. This setup facilitates clean code and resource management, as gRPC’s specifications provide consistency across platforms, and the protocol allows clients to set request deadlines, promoting efficient resource management. Clarifai's API is polyglot, supporting multiple languages, and automatically generates RESTful JSON Web APIs from gRPC services, thus accommodating both gRPC and JSON interfaces without duplicating efforts. gRPC's bi-directional streaming capabilities support real-time communication, vital for many machine learning applications, although not all streaming functionalities are currently utilized. While the gRPC channel is recommended for most use cases, the HTTP+JSON channel offers familiarity for developers and better browser support, albeit at the cost of performance, given that gRPC messages are more compact but not human-readable.