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
- Support for lightweight nets via accelerated depthwise convolutions (BVLC#5665) and shufflenet layer (https://github.com/farmingyard/ShuffleNet).
- Support for image segmentation, via PSPNet, U-Net, SegNet, etc...
- Support for Squeeze & Excitation Nets (https://github.com/hujie-frank/SENet).
- Support for SoftNMS with SSD (https://arxiv.org/abs/1704.04503)
- Support for Focal-Loss with SSD (https://arxiv.org/abs/1708.02002)
- Support for AMSGrad (fix to Adam optimizer, https://openreview.net/forum?id=ryQu7f-RZ)
- Support for Connectionist Temporal Classification (CTC) from https://github.com/baidu-research/warp-ctc and https://github.com/xmfbit/warpctc-caffe
- Support for SGDR (http://arxiv.org/abs/1608.03983)
- Support for RefineDet object detection (https://arxiv.org/abs/1711.06897)
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.