/efficientnet_tf_pytorch_caffe

Primary LanguagePython

EfficientNet for BVLC Caffe

This is a conversion tool for EfficientNet from TensorFlow to BVLC Caffe, based on efficientnet_tf_pytorch_caffe by xxworkspace. This fork uses only BVLC standard layers, instead of broadcast_Mull and Swish layers.

Install packages

Install pip packages

pip install tensorflow torch torchvision --user
pip3 install tensorflow torch torchvision --user

Install BVLC Caffe

sudo apt update
sudo apt install caffe-cpu

Convert models

Preprocessing Architecture Download file
Standard (ckpts) efficientnet-b[1-5] efficientnet-b[1-5]
AutoAugment (ckptsaug) efficientnet-b[1-7] efficientnet-b[1-7]
RandAugment (randaug) efficientnet-b[57] efficientnet-b[57]-randaug
AdvProp (advprop) efficientnet-b[0-8] efficientnet-b[0-8]

TensorFlow to PyTorch Conversion

Download tf-model

pushd pretrained_tensorflow
download.sh ckptsaug efficientnet-b0
popd

Convert tf-model to pytorch-model

mkdir -p pretrained_pytorch/ckptsaug
pushd convert_tf_pt
python3 convert_params_tf_pytorch.py \
 --model_name efficientnet-b0 \
 --tf_checkpoint ../pretrained_tensorflow/ckptsaug/efficientnet-b0/ \
 --output_file ../pretrained_pytorch/ckptsaug/efficientnet-b0.pth
popd

Test PyTorch model

pushd convert_tf_pt
python3 test_pytorch.py ckptsaug efficientnet-b0
popd

PyTorch to Caffe

Convert pytorch-model to caffe .prototxt

mkdir -p caffemodel/ckptsaug
pushd convert_tf_pt
python3 pytorch2caffe.py ckptsaug efficientnet-b0
popd

Convert pytorch-model to caffe .caffemodel

pushd convert_tf_pt
python3 pytorch2caffe_model.py ckptsaug efficientnet-b0
popd

Test

pushd convert_tf_pt
python3 test_caffe.py ckptsaug efficientnet-b0
popd

data process RGB

  • mean = [0.485, 0.456, 0.406] x 255
  • std = [0.229, 0.224, 0.225] x 255

Correspondence table of conversion

Input size Tensorflow layer Caffe layers
Odd conv2d (kernel_size: 2, stride: 2, pad: same) Convolution (kernel_size: 2, stride: 2, pad: 1)
Even conv2d (kernel_size: 2, stride: 2, pad: same) Convolution (kernel_size: 2, stride: 2, pad: 2) +
Convolution (kernel_size: 2, stride: 2, pad: 1) +
Crop (offset: 1)
---------- depthwise_conv2d Convolution (group=num_output)
---------- swish Sigmoid + Eltwise (operation: PROD)
---------- broadcast_mul Flatten + Scale (with 2 bottom blobs)

Result of conversion

efficientnet-b0 on Netscope

https://kentanabe.github.io/netscope/#/preset/efficientnet-b0

efficientnet-b0.block1