microsoft/NeuronBlocks

Based on the README, here's a technical summary: Models are defined entirely through JSON configuration files that compose reusable neural network blocks from Block Zoo, eliminating boilerplate code for training and inference pipelines. Built on PyTorch (0.4.1+), it supports distributed training across CPU/GPU and multiple platforms (Linux/Windows, Philly, PAI) with a built-in model visualizer for architecture validation. Covers diverse NLP tasks including sentence classification, QA matching, slot tagging, and machine reading comprehension with pre-configured Model Zoo examples.

1,454 stars. No commits in the last 6 months.