Kenning aims towards providing modular, unified execution blocks for deep neural network training, optimization, and runtime frameworks.

These blocks enable:

• dataset management
• model data processing (normalization and output processing) and training
• model optimization using such techniques as quantization, pruning, or clustering
• compilation for a given target hardware
• model evaluation and benchmarking
• model quality and performance summaries
• benchmark comparisons for various models, optimizations, accelerators
• running models using efficient runtimes on target devices

Kenning’s AI toolkit lets you:

• automatically optimize pipelines
• automatically select frameworks
• seamlessly connect optimization implementations from different frameworks without worrying about converting models from one framework to another
• prototype and develop advanced applications utilizing DNNs
• reproducibly run and evaluate optimization experiments with full descriptions in JSON format
• compare optimized models performance-wise and quality-wise and generate nice reports in Markdown or HTML with interactive plots

Kenning's reports on model performance and quality, as well as comparison of models yield both textual descriptions as well as visualizations of:

• inference time, memory, CPU and accelerator usage - mean values, as well as their changes over time
• task-specific metrics, such as accuracy, precision, recall, G-Mean or mAP
• detection and segmentation recall-precision curves
• memory accesses, instruction count - metrics obtainable in simulated environments with Renode