**We have developed a set of tools for converting caffemodel to onnx model to facilitate the deployment of algorithms on mobile platforms.
**However, due to the company secrets involved, we can only provide compiled executable files.
**And provide a script to perform forward reasoning on the onnx model and the original caffe model to compare whether the converted results are normal.
| Input | LRN | Deconvolution |
|---|---|---|
| Input | InnerProduct | Interp |
| VideoData | Reshape | Split |
| Convolution | Transpose | Slice |
| Convolution3D | Gemm | ShuffleChannel |
| ConvolutionDepthwise | DropOut | Axpy |
| DepthwiseConvolution | Concat | Crop |
| BatchNorm | Swish | Power |
| Normalize | Sigmoid | Im2Col |
| BN | Mish | Transpose |
| Scale | BroadcastMul | Permute |
| ReLU | Eltwise | Lstm |
| PReLU | Flatten | LSTM |
| Pooling | MaxUnpool | Reverse |
| Pooling3D | Upsample | Reorg |
| Softmax | Uppooling | SpatialProduct |
| AbsVal | Threshold | Reduction |
Congratulations, if you simply convert the caffe model to the onnx model, you don't need to configure any environment. Just run the executable file we provide under the ubuntu system.
git clone https://github.com/xncaffe/caffe_convert_onnx.git
cd caffe_convert_onnx/cmd
./convert_main --prototxt ../examples/inference/models/caffe/mobilenet_v1/deploy.prototxt \
--caffemodel ../examples/inference/models/caffe/mobilenet_v1/deploy.caffemodel \
--out ../examples/inference/models/onnx/mobilenet_v1.onnxIf you are in China, the clone project is slow and often interrupted, you can consider using the download mirror gitclone.com, refer to the command:
git clone https://gitclone.com/github.com/xncaffe/caffe_convert_onnx.gitOr get the conversion executable tool convert_main from the URL https://download.csdn.net/download/xunan003/87659000?spm=1001.2014.3001.5503
You can also use -h for help
./convert_main -h
Found no caffe root, please set caffe root in env param! If transplanting an onnx model, please ignore this warning.
usage: convert_main [-h] -p PROTOTXT [-c CAFFEMODEL] [-o OUT]
optional arguments:
-h, --help show this help message and exit
-p PROTOTXT, --prototxt PROTOTXT
deploy.prototxt path
-c CAFFEMODEL, --caffemodel CAFFEMODEL
deploy.caffemodel path
-o OUT, --out OUT onnx model output path
If you need to use the forward reasoning program we provide to infer the caffe model and onnx model?
**You need to configure some environments to support their normal operation.
-
First you need to install a working caffe framework and configure its python interface.
-
The python version >= 3.7, we used python 3.7.13.
-
Download the example caffemodel from address https://download.csdn.net/download/xunan003/87658946?spm=1001.2014.3001.5503, And decompress examples.zip and place it in the caffe_convert_onnx project directory.
-
Install dependencies according to the provided requirements.txt.
pip install -r requirements.txt
-
Follow the instructions below to get help on using the inference caffe and onnx models
python caffe_inference.py -h usage: caffe_inference.py [-h] [-r CAFFE_ROOT] [-p PROTOTXT] [-c CAFFEMODEL] [-o OUTPUT] [-v MEAN_VALUES [MEAN_VALUES ...]] [-s SCALE_VALUES [SCALE_VALUES ...]] [-i INPUT_DIR] [-t INPUT_TYPE] [--AllTensor] optional arguments: -h, --help show this help message and exit -r CAFFE_ROOT, --caffe_root CAFFE_ROOT caffe root -p PROTOTXT, --prototxt PROTOTXT caffemodel path -c CAFFEMODEL, --caffemodel CAFFEMODEL caffemodel path -o OUTPUT, --output OUTPUT onnx inference output save path -v MEAN_VALUES [MEAN_VALUES ...], --mean_values MEAN_VALUES [MEAN_VALUES ...] pre-processing mean values -s SCALE_VALUES [SCALE_VALUES ...], --scale_values SCALE_VALUES [SCALE_VALUES ...] pre-processing scale values -i INPUT_DIR, --input_dir INPUT_DIR support path, image and tensor, tensor is txt or bin! -t INPUT_TYPE, --input_type INPUT_TYPE pre-processing network support RGB, BGR and Gray! --AllTensor dump all node output tensor!python onnx_inference.py -h usage: onnx_inference.py [-h] [-m ONNXMODEL] [-o OUTPUT] [-v MEAN_VALUES [MEAN_VALUES ...]] [-s SCALE_VALUES [SCALE_VALUES ...]] [-i INPUT_DIR] [-t INPUT_TYPE] [--AllTensor] optional arguments: -h, --help show this help message and exit -m ONNXMODEL, --onnxmodel ONNXMODEL onnx model path -o OUTPUT, --output OUTPUT onnx inference output save path -v MEAN_VALUES [MEAN_VALUES ...], --mean_values MEAN_VALUES [MEAN_VALUES ...] pre-processing mean values -s SCALE_VALUES [SCALE_VALUES ...], --scale_values SCALE_VALUES [SCALE_VALUES ...] pre-processing scale values -i INPUT_DIR, --input_dir INPUT_DIR support path, image and tensor, tensor is txt or bin! -t INPUT_TYPE, --input_type INPUT_TYPE pre-processing network support RGB, BGR and Gray! --AllTensor dump all node output tensor! -
We explain the parameters of caffe and onnx reasoning as follows.
caffe_root -> This is what you have to configure. You need to install the caffe package and compile its python interface, and then configure it. For example, --caffe_root=/home/caffe/python
mean_values -> This is the mean parameter necessary for pre-processing, the default is [0, 0, 0], which strictly corresponds to the format of your --input_type configuration. If --input_type=RGB, the corresponding relationship is r_mean_value=mean_values[0], g_mean_value=mean_values[1], b_mean_value=mean_values[2]. If --input_type=BGR then b_mean_value=mean_values[0], r_mean_value=mean_values[2]. If --input_type=Gray, y_mean_value=mean_values[0].
scale_values -> Similar to mean_values, it is a parameter for pre-processing normalization, and the default value is [1, 1, 1]. Same as mean_values, its order strictly corresponds to --input_type.
input_dir -> Network input file, which can be a path or a specific file. The single-input network is a file, and the multi-input network must be a folder.
input_type -> As explained in mean_values, support RGB\BGR and Gray input, configure according to the actual input of the network. Default is RGB.
AllTensor -> Turn it on if you want to spit out the output of all layers.
Except that caffe_root does not need to be configured, other parameters are consistent with caffe.
Note: Our pre-processing formula is y=(x-mean_values)*scale_values.
-
Example of forward inference using two modelsexamples.
python caffe_inference.py --caffe_root ../caffe/python/ \ --onnxmodel ./examples/inference/models/onnx/twoInputNetDemo.onnx \ --output ./dump_rslts/ \ --mean_values 127.5 127.5 127.5 \ --scale_values 0.0078 0.0078 0.0078 \ --input_dir ./data/multi_input/image$ \ --input_type BGR ### Caffe multi input network image input ###python caffe_inference.py --caffe_root ../caffe/python/ \ --onnxmodel ./examples/inference/models/onnx/mobilenet_v1.onnx \ --output ./dump_rslts/ \ --input_dir ./data/1x3x224x224_float32.txt \ ### Caffe single input network txt input ###python caffe_inference.py --caffe_root ../caffe/python/ \ --onnxmodel ./examples/inference/models/onnx/mobilenet_v1.onnx \ --output ./dump_rslts/ \ --input_dir ./data/1x3x224x224_float32.bin \ ### Caffe single input network bin input ###python caffe_inference.py --caffe_root ../caffe/python/ \ --onnxmodel ./examples/inference/models/onnx/mobilenet_v1.onnx \ --output ./dump_rslts/ \ --mean_values 127.5 127.5 127.5 \ --scale_values 0.0078 0.0078 0.0078 \ --input_dir ./data/1.jpg \ --input_type BGR ### Caffe single input network image input ###python onnx_inference.py --onnxmodel ./examples/inference/models/onnx/mobilenet_v1.onnx \ --output ./dump_rslts/ \ --mean_values 127.5 127.5 127.5 \ --scale_values 0.0078 0.0078 0.0078 \ --input_dir ./data/1.jpg \ --input_type BGR ### Onnx single input network image input ###
For other tests, please operate similarly.
Note that multi-input networks only support the case where all inputs are the same set of pre-processing parameters.