brocolli

a pytorch to caffe model converter, it learns from MMdnn && PytorchConverter. MMdnn only provides tools for pytorch to onnx conversion, but our tool provides direct conversion from pytorch to caffe

How to

run local

git clone https://github.com/inisis/caffe (only for valiation if u don't need verify your results, skip this)

cd caffe && (revise your dependency to caffe 3) && make pycaffe

pip3 install torch==0.4.0 torchvision==0.2.0

export PYTHONPATH=$PYTHONPATH:/your/path/to/brocolli/

put your pytorch model in tests/pytorch_model folder

python3 pytorch_model_converter.py

python3 ssd_layer.py example.json pytorch_model/best.pth.prototxt new.prototxt

python3 caffe_test.py

python3 pytorch_test.py

run container

docker pull yaphets4desmond/pytorch_converter_stable

Notice

Curently supported layers

Conv

PRelu

MaxPooling

Sigmoid

BatchNormalization

Relu

Add

AvgPool

Flatten

FullyConnected

Dropout

Softmax

Upsample

Permute

Concat

Curently supported network

SSD

ResNet

Contact