caffe

Its C++ core with Python and MATLAB interfaces, combined with GPU acceleration, enabled the training of image classification models like AlexNet and GoogLeNet at speeds that were groundbreaking for the time.

Caffe's prototext-based model definition format allowed researchers to define network architectures declaratively, before the era of imperative PyTorch-style model building.

Caffe's design prioritized speed and production deployment over flexibility: its static graph architecture and efficient memory management made it a preferred framework for deploying CNN models in production computer vision applications, particularly on GPU-constrained hardware.

The Berkeley-maintained Model Zoo provided pretrained weights for common architectures, making transfer learning accessible before the concept became widely standardized.

Caffe2, developed later by Facebook, extended Caffe's production deployment strengths and was eventually merged into PyTorch as its mobile and production execution backend.

While largely superseded by PyTorch and TensorFlow for new model development, Caffe retains historical significance as the framework that enabled the computer vision research renaissance in the mid-2010s.

Legacy computer vision systems deployed in the 2015-2018 period at industrial scaleproduction image classification, face detection, content moderationmay still run Caffe models.