This is a chainer implementation of FC-DenseNets (also named Tiramisu103)
The network architecture was described in The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation. However I found some contradiction between the content of the paper and their official implementation on Theano.
This chainer implementation is basically a "translation" of the official theano implementation (it means, I chose to believe the official code rather than the paper when dealing with inconsistency), while having some uncertainty:
-
Original code use 3x3 deconvolution to do upsampling, but I don't know how to double the resolution size of the feature maps precisely by 3x3 deconvolution in chainer. So I use 2x2 deconv instead. https://github.com/haqishen/chainer-FC-DenseNet-Tiramisu/blob/master/Tiramisu.py#L61
-
The author said that there are 1088 channels in the first upsampling dense block (https://github.com/SimJeg/FC-DenseNet/blob/master/FC-DenseNet.py#L98-L108) I'm not sure about that, as I can only see 848 channels there.