build

• pip install tensorflow torch torchvision
• pip3 install tensorflow torch torchvision
• git clone https://github.com/xxworkspace/caffe.git
  • Then build caffe width cpu-only

TensorFlow to PyTorch Conversion

• Download tf-model
  • cd pretrained_tensorflow
  • bash download.sh efficientnet-b*
• Convert tf-model to pytorch-model
  • cd convert_tf_pt
  • python3 convert_params_tf_pytorch.py --model_name efficientnet-b0 --tf_checkpoint ../pretrained_tensorflow/efficientnet-b0/ --output_file ../pretrained_pytorch/efficientnet-b0.pth
    • Using python3 convert_params_tf_pytorch.py -h
• Test
  • python test_pytorch.py efficientnet-b0*

PyTorch to Caffe

• Convert pytorch-model to caffe .prototxt
  • python pytorch2caffe.py efficientnet-b*
• Convert pytorch-model to caffe .caffemodel
  • python pytorch2caffe_model.py efficientnet-b*
• Test
  • pytorch test_caffe.py efficientnet-b*

data process RGB

mean = [0.485, 0.456, 0.406] std = [0.229, 0.224, 0.225]