PyTorch2Caffe

Require Pytorch < 0.4

Ported from pytorch-caffe-darknet-convert.

Add support for

• Dilated Convolution Layer

• Concat Layer

• Upsampling (converted to Deconvolution with bilinear initialization)

• Eltwise Product

• Sigmoid Layer

# We can obtain almost the same output from caffe except Upsampling# for inception_v3: # diff between pytorch and caffe: min: 0.0, max: 1.76429748535e-05, mean: 2.14079022953e-06# see more in demo.pyimport torchfrom torch.autograd import Variableimport torchvisionimport osfrom pytorch2caffe import pytorch2caffe, plot_graph

m = torchvision.models.inception_v3(pretrained=True, transform_input=False)
m.eval()print(m)

input_var = Variable(torch.rand(1, 3, 299, 299))
output_var = m(input_var)

output_dir = 'demo'# plot graph to pngplot_graph(output_var, os.path.join(output_dir, 'inception_v3.dot'))

pytorch2caffe(input_var, output_var, 
              os.path.join(output_dir, 'inception_v3-pytorch2caffe.prototxt'),
              os.path.join(output_dir, 'inception_v3-pytorch2caffe.caffemodel'))