BackPACK library does not support some of the modules in the model, specifically **GroupNorm**.
For this should I be creating custom extensions for the unsupported modules?
Logs when training the model.
env/lib/python3.11/site-packages/backpack/extensions/backprop_extension.py:106: UserWarning: Extension saving to grad_batch does not have an extension for Module <class 'networks.UNet'> although the module has parameters
warnings.warn(
env/lib/python3.11/site-packages/backpack/extensions/backprop_extension.py:106: UserWarning: Extension saving to grad_batch does not have an extension for Module <class 'torch.nn.modules.normalization.GroupNorm'> although the module has parameters
I assume from your logs that you would like to support BatchGrad for nn.GroupNorm.
You can follow the instructions here to achieve that. A PR adding this to BackPACK would be really cool, too :)
I am re-implementing the enhancement of DP-SGD through the random sparsification of gradients on my UNet Model.
Here is a Debug info on extending the Model
extend(model)
.BackPACK library does not support some of the modules in the model, specifically
**GroupNorm**
.For this should I be creating custom extensions for the unsupported modules?
Logs when training the model.
Thanks for the help :)