Felix Dangel

Hi, thanks for your feature request. we have an [example](https://docs.backpack.pt/en/master/use_cases/example_custom_module.html#sphx-glr-use-cases-example-custom-module-py) how to add new parameterized layers to first-order extensions. It's a good starting point. Since `nn.Upsample` has no parameters, you...

Hi, I briefly glanced over the linked documentation. Two questions for your specific use case, (i) do you use `GaussianNLLLoss` with `full=False` (and do I understand the doc correctly that...

The bug still persists in `torch==1.9.1`. @schaefertim could you update the [skip condition](https://github.com/tjs111/backpack/blob/b9b11af594e72408428a48b11de96b59ac57f183/test/utils/skip_test.py#L13-L29) in the tests?

Fixed in `torch=1.10.0` ([#60630](https://github.com/pytorch/pytorch/pull/60630)).

Hi, thanks for your question. As many of Cockpit's quantities rely on BackPACK, your model architecture must be supported to get the full view. You can take a look at...

Sounds good. By the way, BackPACK has basic support for ResNets from version `1.4.0` on. Check out [this example](https://docs.backpack.pt/en/1.4.0/use_cases/example_resnet_all_in_one.html#any-resnet-with-backpack-s-converter) to see how it works.

Hi, I think we need more information to reproduce your issue. The error message you obtain does not seem to be related to your loss being non-standard but rather a...

Hi @thomasahle, I'm not sure I fully understand your question. Individual gradients don't exist in a network with BN layer in train mode, because the individual losses depend on all...

Hi Thomas, > Right now the non batched `grad` is only computed whenever `batch_l2` is computed. That seems odd to me because BackPACK does not intervene into PyTorch's gradient computation....

Hi, thanks for reporting this problem (related to https://github.com/fKunstner/backpack-discuss/issues/65). I confirm the mentioned example does not work. Check out [this](https://docs.backpack.pt/en/master/basic_usage/example_all_in_one.html) example which should work.