jzbontar
commented
8 years ago There is no data processing involved. As we wrote in the paper: a convolutional layer can be applied to inputs of arbitrary sizes. The only complication is with fully-connected layers, which have to be replaced with convolutional layers with 1x1 kernels.
To see how this is done in the code, see:
- https://github.com/jzbontar/mc-cnn/blob/master/main.lua#L661
- https://github.com/jzbontar/mc-cnn/blob/master/main.lua#L678
The network net_tr is the network used for training---it has fully-connected layers and expects 9x9 images on input; while net_te is the network used at test time, where the fully-connected layers are replaced with convolutions.
I’m confused about a question.In your paper, we can see your three implementation details kept the running time manageable.While you training the network with 9*9 patches, if you want to compute for all pixels in a single forward pass by propagating full-resolution images, I guess there must be a part of data-processing, can you give me detailed descriptions about this part? Hope your early reply.