MNNConvert

编译模型转换工具(gcc>=4.9)

首先需要安装protobuf(3.0以上)

  1. # MacOS
  2. brew install protobuf

其它平台请参考官方安装步骤

  1. cd MNN/tools/converter
  2. ./generate_schema.sh
  3. mkdir build
  4. cd build && cmake .. && make -j4
  6. # 或者直接执行脚本
  7. ./build_tool.sh

模型转换的使用

  1. Usage:
  2.   MNNConvert [OPTION...]
  4.   -h, --help            Convert Other Model Format To MNN Model
  6.   -v, --version         show current version
  7.   -f, --framework arg   model type, ex: [TF,CAFFE,ONNX,TFLITE,MNN]
  8.       --modelFile arg   tensorflow Pb or caffeModel, ex: *.pb,*caffemodel
  9.       --prototxt arg    only used for caffe, ex: *.prototxt
  10.       --MNNModel arg    MNN model, ex: *.mnn
  11.       --benchmarkModel  Do NOT save big size data, such as Conv's weight,BN's
  12.                         gamma,beta,mean and variance etc. Only used to test
  13.                         the cost of the model
  14.       --bizCode arg     MNN Model Flag, ex: MNN
  15.       --debug           Enable debugging mode.

说明: 选项benchmarkModel将模型中例如卷积的weight，BN的mean，var等参数移除，减小转换后模型文件大小，在运行时随机初始化参数，以方便测试模型的性能。

tensorflow/ONNX/tflite

  1. ./MNNConvert -f TF/ONNX/TFLITE --modelFile XXX.pb/XXX.onnx/XXX.tflite --MNNModel XXX.XX --bizCode XXX

三个选项是必须的！ 例如:

  1. ./MNNConvert -f TF --modelFile path/to/mobilenetv1.pb --MNNModel model.mnn --bizCode MNN

caffe

  1. ./MNNConvert -f CAFFE --modelFile XXX.caffemodel --prototxt XXX.prototxt --MNNModel XXX.XX --bizCode XXX

四个选项是必须的！ 例如:

  1. ./MNNConvert -f CAFFE --modelFile path/to/mobilenetv1.caffemodel --prototxt path/to/mobilenetv1.prototxt --MNNModel model.mnn --bizCode MNN

MNN

  1. ./MNNConvert -f MNN --modelFile XXX.mnn --MNNModel XXX.XX --bizCode XXX

查看版本号

  1. ./MNNConvert --version

MNNDump2Json

将MNN模型bin文件 dump 成可读的类json格式文件，以方便对比原始模型参数

Pytorch 模型转换

  1. import torch
  2. import torchvision
  4. dummy_input = torch.randn(10, 3, 224, 224, device='cuda')
  5. model = torchvision.models.alexnet(pretrained=True).cuda()
  7. # Providing input and output names sets the display names for values
  8. # within the model's graph. Setting these does not change the semantics
  9. # of the graph; it is only for readability.
  10. #
  11. # The inputs to the network consist of the flat list of inputs (i.e.
  12. # the values you would pass to the forward() method) followed by the
  13. # flat list of parameters. You can partially specify names, i.e. provide
  14. # a list here shorter than the number of inputs to the model, and we will
  15. # only set that subset of names, starting from the beginning.
  16. input_names = [ "actual_input_1" ] + [ "learned_%d" % i for i in range(16) ]
  17. output_names = [ "output1" ]
  19. torch.onnx.export(model, dummy_input, "alexnet.onnx", verbose=True, input_names=input_names, output_names=output_names)
  • 将 Onnx 模型文件转成 MNN 模型
  1. ./MNNConvert -f ONNX --modelFile alexnet.onnx --MNNModel alexnet.mnn --bizCode MNN