After converting my PyTorch model to Caffe2, my model gives slightly different result. It is due to the slight numerical difference in the implementation of convolution layer. Hopefully, there is a way for PyTorch to use Caffe2’s version of convolution layer.