This is a slightly modified version of Caffe as used by the Deep Learning API & server Deepdetect. The repository is kept up to date with the original Caffe master branch.
Improvements and new features include:
- Switch from
LOG(FATAL)error toCaffeErrorExceptionthrown on every recoverable errors. This allows the safe use of Caffe as a C++ library from external applications, and in production - Various fixes, including ability to run the exact same job in parallel
- Makefile fixes with default build supporting all NVIDIA architectures
- Sparse inputs and CPU/GPU computations
- Support for class weights applied to Softmax loss, useful for training over imbalanced datasets
- SSD: Single Shot MultiBox Detector for object detection in images
While this is intended to be used with DeepDetect, this is a great alternative to the original Caffe if you'd like to avoid uncaptured errors, train from text or sparse data, need built-in image detection.