Skip to content

Latest commit

 

History

History
executable file
·
16 lines (15 loc) · 1.13 KB

README.md

File metadata and controls

executable file
·
16 lines (15 loc) · 1.13 KB
Raw
  1. symbol转prototxt: json2prototxt.py, 转出后手动把不支持的地方接上
  2. 权重转换：mxnet2caffe.py
  3. 如果想两步合一：mxnet2caffe_all_in_one.py，如果有不支持的层就不能2步合一了
  4. 如果需要调整输入顺序：caffe_swap_channel_toolkit.py
  5. 如果需要inplace操作：inplace_conv_block.py, 注意转换后可能需要手动微调
  6. insert_mean_scale.py, 将均值和方差用Scale层的形式写入模型文件
  7. mxnet2caffe_all_in_one_experimental.py: 与步骤三功能一致，区别是自动跳过不支持的层

English Guidelines:

  1. symbol file convert to prototxt: json2prototxt.py, if there are some layers skipped, modify the prototxt file mannualy
  2. convert params: mxnet2caffe.py
  3. 1,2 steps in one: mxnet2caffe_all_in_one.py, however, if there are some layers skipped, we cannot use it.
  4. swap input channel(bgr->rgb or vice versa): caffe_swap_channel_toolkit.py
  5. inplace op: inplace_conv_block.py
  6. use Scale layer to process data mean and std: insert_mean_scale.py
  7. mxnet2caffe_all_in_one_experimental.py: same with step 3 but skip the unsupported layers automaticaly