README.md

pytorch2caffe

A package to convert pytorch model to caffe model.

Note

• This package supports F.interpolate with mode='bilinear' argument and scale_factor argument. It depends on caffe with Upsample layer as follows:

# argument height and width
layer{
    name:"Upsample_nearest"
    type:"Upsample"
    bottom:"Conv2d_87" #Blob Conv2d_87's shape is [1,16,32,32]
    top:"Upsample_nearest" #Blob Upsample_nearest's shape is [1,16,HEIGHT,WIDTH]
    upsample_param{
        mode: NEAREST # or BILINEAR
        height: HEIGHT
        width: WIDTH
    }
}
# argument height_scale and width_scale
layer{
    name:"Upsample_nearest"
    type:"Upsample"
    bottom:"Conv2d_87" #Blob Conv2d_87's shape is [1,16,32,32]
    top:"Upsample_nearest" #Blob Upsample_nearest's shape is [1,16,32*HEIGHT_SCALE,32*WIDTH_SCALE]
    upsample_param{
        mode: NEAREST # or BILINEAR
        height_scale: HEIGHT_SCALE
        width_scale: WIDTH_SCALE
    }
}

See jnulzl/caffe_plus

Installation

git clone https://github.com/wgggSg/pytorch2caffe
cd pytorch2caffe
python3 setup.py install

Tutorials

Note: import torch first then import pytorch2caffe

import torch
from torchvision.models import resnet
from pytorch2caffe import pytorch2caffe

name='resnet18'
resnet18=resnet.resnet18()
resnet18.eval()
dummy_input=torch.ones([1,3,224,224])
pytorch2caffe.trans_net(resnet18, dummy_input, name)
pytorch2caffe.save_prototxt('{}.prototxt'.format(name))
pytorch2caffe.save_caffemodel('{}.caffemodel'.format(name))