nikkkkhil / model-exchange

model-exchange

model-exchange is a model conversion and visualization tool to help users inter-operate among different deep learning frameworks. Convert models between Keras, MXNet, PyTorch, Caffe and Tensorflow.

Requirments

• tensorflow==1.13.0
• pycaffe
• keras==2.4.1
• pytorch==0.4.1
• mxnet==1.0.0
• protobuf==3.6.1

How to deploy visualization on Web

If you want to access the deployed web page from an external network, you should first modify the host and port for the ./visualization/app.py file.

You can deploy it on Web by running:

python app.py

Exmaples

MXNet <-> IR

cd scripts/

• Convert MXNet to IR

  CUDA_VISIBLE_DEVICES=0 python convertToIR.py -s mxnet -d outname -n path/to/network -w path/to/weight/file

• Convert IR to MXNet

  CUDA_VISIBLE_DEVICES=0 python IRtoModel.py -f mxnet -d path/to/save/the/destination/model -n path/to/IR/network/structure/file -w path/to/IR/weight/file

Caffe <-> IR

cd scripts/

• Convert caffe to IR

  CUDA_VISIBLE_DEVICES=0 python convertToIR.py -s caffe -d outname -n path/to/network -w path/to/weight/file

• Convert IR to Caffe

  CUDA_VISIBLE_DEVICES=0 python IRtoModel.py -f caffe -d path/to/save/the/destination/model -n path/to/IR/network/structure/file -w path/to/IR/weight/file

  The resulting conversion failure is dependent on the protobuf version inconsistency.

PyTorch <-> IR

cd scripts/

• Convert pytorch to IR

  CUDA_VISIBLE_DEVICES=0 python convertToIR.py -s pytorch -d outname -n path/to/network -w path/to/weight/file

• Convert IR to pytorch

  CUDA_VISIBLE_DEVICES=0 python IRtoModel.py -f pytorch -d path/to/save/the/destination/model -n path/to/IR/network/structure/file -w path/to/IR/weight/file

Keras -> IR

cd scripts/

• Convert keras to IR

  CUDA_VISIBLE_DEVICES=0 python convertToIR.py -s keras -d outname -n path/to/network -w /path/to/weight/file

Tensorflow <-> IR

Convert tf to IR

• Download resnet_v1_101 data

  $ wget http://download.tensorflow.org/models/resnet_v1_101_2016_08_28.tar.gz
  $ tar -xvf resnet_v1_101_2016_08_28.tar.gz

• Extract tf model files Refer the common/tensorflow/extractor.py to implement your own model extract code.

  CUDA_VISIBLE_DEVICES=0 python ./common/tensorflow/extractor.py -n resnet_v1_101 -p path/to/network/file -o path/to/outdir

• Convert tf to IR

  CUDA_VISIBLE_DEVICES=0 python convertToIR.py -s tf -d outname -n path/to/network --dstNodeName Squeeze -w path/to/weight/file

Convert IR to tf

• Convert tf to IR

  cd scripts/
  CUDA_VISIBLE_DEVICES=0 python ./scripts/IRtoModel.py --phase test/train -f tf -d path/to/save/the/destination/model -n path/to/IR/network/structure/file -w path/to/IR/weight/file

Acknowledgements

Thanks to Microsoft, the initial code of MXNet -> IR converting is references to his project MMdnn.