lixiaolei1982 / keras-implementation-of-u-net-r2u-net-attention-u-net-attention-r2u-net.- Goto Github PK

Model has randomness issue every time we run it, results are not reproducible.
But yeah after kernel restart and notebook restart same results can be obtained.
But it failed for 5-Fold case.

is the R2UNet corrct? i check the pytorch R2Unet which have 1.2k stars, and i print his model,the params are 39,091,521

and i use model summary in your model, it shows 95,986,177

The given model does not use Recurrent Convolutional Layers, which are described in the R2-Unet paper "Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation"
This can be seen in the function rec_res_block, where there are no weights sharing at all. Increasing the number of iterations in the inner loop should not introduce new convolutional layer, but only enlarge the receptive field of one convolutional layer. This can be seen in the paper as Wk_f and Wk_r which do not have t as their parameter (which means that these weights are the same set of weights through t0, t1...).