Keras implementation of the paper "3D MRI brain tumor segmentation using autoencoder regularization" by Myronenko A. (https://arxiv.org/abs/1810.11654).
The Dense layers Dec_VAE_VDraw_Mean and Dec_VAE_VDraw_Var should just be Identity layers taking the first and last 128 outputs of Dec_VAE_VD_Dense, as per section 3.3 of the paper:
"Starting from the encoder endpoint output, we first reduce the input to a low
dimensional space of 256 (128 to represent mean, and 128 to represent std).
Then, a sample is drawn from the Gaussian distribution with the given mean
and std (...)".
Right now there is an additional linear 256x128 mapping between the low-dim. space and both mu and sigma.
The
DenselayersDec_VAE_VDraw_MeanandDec_VAE_VDraw_Varshould just be Identity layers taking the first and last 128 outputs ofDec_VAE_VD_Dense, as per section 3.3 of the paper:"Starting from the encoder endpoint output, we first reduce the input to a low dimensional space of 256 (128 to represent mean, and 128 to represent std). Then, a sample is drawn from the Gaussian distribution with the given mean and std (...)".
Right now there is an additional linear 256x128 mapping between the low-dim. space and both mu and sigma.