Theano @ Dragon
计算图原语
Caffe @ Dragon
子集问题
- Blob →Tensor
- Layer → Op
- Net → Graph
- Solve ->Optimizer
2000行复现Caffe
Caffe 与 PyCaffe
Caffe: C++ 接口笨重,编译复杂,扩展库少;
PyCaffe: PyCaffe接口灵活,无穷无尽的Python科学计算库。
Dragon-Caffe
共享Dragon的符号仓库与分布式优化器;
比原生Caffe节省50%~75%的显存。
Caffe-Dragon示例
1 | import dragon.vm.caffe as caffe |
源码阅读
FCN图像分割:https://github.com/shelhamer/fen.berkeleyvising.org
Faster-RCNN目标检测:https://github.com/rbgirshick/py-faster-rcnn
Tensorflow @ Dragon
子集问题
- Tenosr’ → Tensor
- Op’ → Op
- Graph’ → Graph
- Session → Function
2000复现TensorFlow
TensorFlow-Dragon 示例
1 | import dragon.m.tensorflow as tf |