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.py
import torch
from torch.autograd import Variable
import torchvision
import os
from 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 png
plot_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'))