search

MIC-DKFZ / nnUNet

Apache License 2.0
5.32k stars 1.63k forks source link

ResNet encoder difference #2312

Open Malitha123 opened 6 days ago

Malitha123 commented 6 days ago

Hi,

Just out of curiosity, what is the fundamental difference between the ResEnc M, ResEnc L and ResEnc XL in architecture wise? Does it change the depth of the network?

GregorKoehler commented 5 days ago

Hi,

unfortunately, I can't give a concise answer to this, but in this file you can follow the algorithm for how the plans are set for a given configuration: https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunetv2/experiment_planning/experiment_planners/residual_unets/residual_encoder_unet_planners.py

Look at the get_plans_for_configuration method and what changes for the M / L / XL configurations (hint: UNet_reference_val_2d / UNet_reference_val_3d).

If that doesn't help, I'd propose creating plans for the individual configurations (M / L / XL) in your case (dataset) and looking at the plans differences.

A rough overview would be: For determining the network topology, this function is used: https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunetv2/experiment_planning/experiment_planners/network_topology.py#L30

The above-mentioned get_plans_for_configuration progressively reduces the patch size to match the target GPU memory estimate and adapts the topology following get_pool_and_conv_props.

I hope this helps!

Malitha123 commented 4 days ago

Thank you for your response. I will take a look