Residual Encoder are much slower to train

[romainVala]

Hello

First thanks for this very nice tool !

I recently give a try with the new preset, and run a training with nnUNetResEncUNetLPlans

Sorry if I miss the information but I could not found benchmark for those model,
on the exact same dataset, Is it expected that the training take 4 time more compare the standard 3D_fullres ? (I have an epoch time around 300 whereas, it takes ony 75 for the 3D_fullres)

I have an other (may be related) question about the model parameter, I was surprise to see very different setting for encoder and decoder : (after running nnUNetPlannerResEncL)

nnunet found the following parameter for ResEncL

"n_blocks_per_stage": [ 1,  3, 4, 6, 6, 6, 6 ],
 "n_conv_per_stage_decoder": [1, 1, 1, 1, 1, 1 ],

whereas for the 3D_fullres, I get something more expected with

"n_blocks_per_stage": [ 2, 2, 2, 2, 2, 2 ],
 "n_conv_per_stage_decoder": [2, 2, 2, 2, 2, ],

is that expected ?

Thanks