BVLC/caffe

Built on modular layer-based architecture with CPU/GPU acceleration via CUDA, Caffe excels at convolutional neural networks for computer vision tasks. Configuration-driven design separates model definitions (prototxt) from training code, enabling rapid experimentation without recompilation. Integrates with multiple backends including OpenCL and Intel optimizations, plus includes a pre-trained model zoo for transfer learning.

34,770 stars. No commits in the last 6 months.