Skip to content

Latest commit

 

History

History
125 lines (101 loc) · 5.27 KB

caffe2tnn_en.md

File metadata and controls

125 lines (101 loc) · 5.27 KB
Raw

Caffe Model to ONNX Model

中文版本

To convert the Caffe model to a TNN model, first convert the Caffe model to an ONNX model, which is then converted to a TNN model.

We use the caffe2onnx tool to convert the Caffe model to ONNX. In the following document, it will briefly introduce how to use caffe2onnx to convert, and then it is recommended to refer to onnx2tnnto convert the ONNX model to TNN.

1. Environment(Mac and Linux)

  • install protobuf(version >= 3.4.0)

Macos:

brew install protobuf

Linux:

For Linux system，we recommend following protobuf's official README.

If you are using the Ubuntu system, use the instructions below to install:

sudo apt-get install libprotobuf-dev protobuf-compiler
  • Install python (version >=3.6)

Macos

brew install python3

centos:

yum install  python3 python3-devel
  • onnx(version == 1.6.0)
pip3 install onnx==1.6.0
  • numpy(version >= 1.17.0)
pip3 install numpy

2. caffe2onnx tool usage

  • cd in tool directory
cd <tnn_root_path>/tools/caffe2onnx/
  • Caffe format conversion

At present, the caffe2onnx tool currently only supports the latest version of caffe format. So before using caffe2onnx, you need to upgrade the old version of the caffe network and model to the latest version. Caffe comes with the tool to convert from old version caffe model and network to a new version.

upgrade_net_proto_text [old prototxt] [new prototxt]
upgrade_net_proto_binary [old caffemodel] [new caffemodel]

The format of the modified input is as follows:

layer {
  name: "data"
  type: "input"
  top: "data"
  input_param { shape: { dim: 1 dim: 3 dim: 224 dim: 224 } }
}
  • caffe2onnx tool usage
python3 convert2onnx.py ./test.prototxt ./test.caffemodel -o ./test.onnx
usage: convert2onnx.py [-h] [-o ONNX_FILE] proto_file caffe_model_file

convert caffe model to onnx

positional arguments:
  proto_file        the path for prototxt file, the file name must end with
                    .prototxt
  caffe_model_file  the path for caffe model file, the file name must end with
                    .caffemodel!

optional arguments:
  -h, --help            show this help message and exit
  -o ONNX_FILE          the path for generate onnx file
  -align                align the onnx model with tnn model
  -input_file in.txt    the input file path which contains the input data for the inference model
  -ref_file   ref.txt   the reference file path which contains the reference data to compare the results

3. caffe2onnx operator support

Number caffe layer onnx operator
1 BatchNorm BatchNormalization
2 BatchNorm + Scale BatchNormalization
3 Concat Concat
4 Convolution Conv
5 ConvolutionDepthwise Conv
6 Deconvolution ConvTranspose
7 DetectionOutput DetectionOutput(customer defination)
8 Dropout Dropout
9 Eltwise Mul/Add/Max
10 Flatten Reshape
11 InnerProduct Reshape + Gemm
12 LRN LRN
13 MaxUnPool MaxUnPool
14 PReLU PRelu
15 Permute Transpose
16 Pooling MaxPool/AveragePool/GlobalMaxPool/GlobalAveragePool
17 PriorBox PriorBox(customer defination)
18 ReLU Relu/LeakyRelu
19 ReLU6 Clip
20 Reshape Reshape
21 Scale Mul + Reshape
22 ShuffleChannel Reshape + Transpose + Reshape
23 Sigmoid Sigmoid
24 Slice Slice
25 Softmax Softmax
26 Upsample Resize