Request: Extension for GroupNorm

[ParthS007]

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).

UNet(
  (encoder1): Sequential(
    (enc1conv1): Conv2d(1, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (enc1norm1): GroupNorm(32, 32, eps=1e-05, affine=True)
    (enc1relu1): ReLU()
    (enc1conv2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (enc1norm2): GroupNorm(32, 32, eps=1e-05, affine=True)
    (enc1relu2): ReLU()
  )
  (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  ....
  
  (bottleneck): Sequential(
    (bottleneckconv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (bottlenecknorm1): GroupNorm(32, 512, eps=1e-05, affine=True)
    (bottleneckrelu1): ReLU()
    (bottleneckconv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (bottlenecknorm2): GroupNorm(32, 512, eps=1e-05, affine=True)
    (bottleneckrelu2): ReLU()
  )
  (upconv4): ConvTranspose2d(512, 256, kernel_size=(2, 2), stride=(2, 2))
  (decoder4): Sequential(
    (dec4conv1): Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (dec4norm1): GroupNorm(32, 256, eps=1e-05, affine=True)
    (dec4relu1): ReLU()
    (dec4conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (dec4norm2): GroupNorm(32, 256, eps=1e-05, affine=True)
    (dec4relu2): ReLU()
  )
  ...
)

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

Thanks for the help :)

[f-dangel]

Hi Parth,

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 :)

Best,
Felix