Query Version

./MNNConvert --version

Model Convert

Parameters

Usage:
  MNNConvert [OPTION...]
  -h, --help                    Convert Other Model Format To MNN Model
  -v, --version                 show current version
  -f, --framework arg           model type, ex: [TF,CAFFE,ONNX,TFLITE,MNN]
      --modelFile arg           tensorflow Pb or caffeModel, ex:
                                *.pb,*caffemodel
      --prototxt arg            only used for caffe, ex: *.prototxt
      --MNNModel arg            MNN model, ex: *.mnn
      --fp16                    save Conv's weight/bias in half_float data
                                type
      --benchmarkModel          Do NOT save big size data, such as Conv's
                                weight,BN's gamma,beta,mean and variance etc.
                                Only used to test the cost of the model
      --bizCode arg             MNN Model Flag, ex: MNN
      --debug                   Enable debugging mode.
      --forTraining             whether or not to save training ops BN and
                                Dropout, default: false
      --weightQuantBits arg     quantize conv/matmul/LSTM float weights to int
                                type, only optimize for model size, optional 2-8 bits,
                                default: 0, which means no weight quant
      --compressionParamsFile arg
                                The path of the compression parameters that
                                stores activation, weight scales and zero
                                points for quantization or information for
                                sparsity.
      --saveStaticModel         save static model with fix shape, default:
                                false
      --inputConfigFile arg     set input config file for static model, ex:
                                ~/config.txt

Note1: The benchmarkModel option removes some parameters from the model to reduce the size of it, such as weight of convolution, mean, var of BN. removed parameters will be initialized randomly in runtime. This will be helpful in performance testing.

Note2: The weightQuantBits option should be used as “—weightQuantBits numBits”, where numBits=2~8. this function is to quantize conv/matmul/LSTM float weights to int type, only optimize for model size, optional 2-8 bits. the weight-quantized model will be decoded to float32 when it is loaded. inference speed is the same with the float32 model, but 4X smaller(8bits), the accuracy is almost the same with its original float32 version.

TensorFlow -> MNN

./MNNConvert -f TF --modelFile XXX.pb --MNNModel XXX.mnn --bizCode biz

TensorFlow Lite -> MNN

./MNNConvert -f TFLITE --modelFile XXX.tflite --MNNModel XXX.mnn --bizCode biz

Caffe -> MNN

./MNNConvert -f CAFFE --modelFile XXX.caffemodel --prototxt XXX.prototxt --MNNModel XXX.mnn --bizCode biz

ONNX -> MNN

./MNNConvert -f ONNX --modelFile XXX.onnx --MNNModel XXX.mnn --bizCode biz

PyTorch -> MNN

1. Convert PyTorch model to ONNX (https://pytorch.org/docs/stable/onnx.html)

   import torch
   import torchvision
   dummy_input = torch.randn(10, 3, 224, 224, device='cuda')
   model = torchvision.models.alexnet(pretrained=True).cuda()
   # Providing input and output names sets the display names for values
   # within the model's graph. Setting these does not change the semantics
   # of the graph; it is only for readability.
   #
   # The inputs to the network consist of the flat list of inputs (i.e.
   # the values you would pass to the forward() method) followed by the
   # flat list of parameters. You can partially specify names, i.e. provide
   # a list here shorter than the number of inputs to the model, and we will
   # only set that subset of names, starting from the beginning.
   input_names = [ "actual_input_1" ] + [ "learned_%d" % i for i in range(16) ]
   output_names = [ "output1" ]
   torch.onnx.export(model, dummy_input, "alexnet.onnx", verbose=True, input_names=input_names, output_names=output_names)

2. Convert ONNX to MNN

   ./MNNConvert -f ONNX --modelFile XXX.onnx --MNNModel XXX.mnn --bizCode biz

Python Tools for Model Conversion

We provided the model conversion tools in our Python toolchain. See here: https://www.yuque.com/mnn/en/usage_in_python