Awesome

PyTorch implementation of Convolutional Neural Fabrics arxiv:1606.02492 There are some minor differences:

• The raw image is first convolved, to obtain #channels feature maps.
• The upsampling is followed by a convolution, and the result is then summed with the other inputs. In the paper, they first sum and then convolve on the result.
• These can be easily changed in the UpSample, DownSample, SameRes class definitions inside neural_fabrics.py. Feel free to implement your own procedure and experiment.

To run on CIFAR-10:

Test set error: 7.2%, with rotation and translation augmented training data.