ensta-u2is-ai / torch-uncertainty Goto Github PK

I have added dropout to standard ResNet in de31452. It could be interesting to add an MCDropout Ensemble baseline. To do this, the following steps could be considered:

• Create a new baseline in torch-uncertainty/torch_uncertainty/baselines/classification/resnet.py in cls.ensemble
• Find a solution to apply dropout in eval mode (potentially change nn.Dropout to F.dropout and replace self.training by true when the version is MCdropout)
• Add tests
• Add a tutorial to explain how to use the baseline
• Add the reference to the documentation
• Add a parameter to enable last-layer dropout

• Add a new routine
• Add a model
• Add NYUv2
• Add MUAD Depth
• Add metrics

The regression routines should allow the user to choose the distribution of liking. For now, we cannot distinguish Laplace and Gaussian distributions as we rely on the number of parameters to choose the corresponding distributions.

• Implement at least the following distributions: Dirac (standard), Gaussian, Laplace, NIG

A potential solution was highlighted in this discussion:

"""
I had the same kind of problem with another project of mine, I would suggest creating a function like this:

def get_distribution(dist_name: str, dist_params: Tensor) -> Distribution:
    ...

I think it could be interesting to use the Distribution class from Pytorch, but it might be too much. A dictionary with the parameters should do it nicely too.
"""

Originally posted by @alafage in #46 (comment)

This metric could be added to the default group in the classification routine.

More information on this dataset on PapersWithCode

This dataset is straightforward to implement. Here is a possible roadmap:

• Add the dataset to the datasets folder. Get inspiration from CIFAR-H and C, for instance.
• Check that the dataset is downloaded correctly & automatically
• Add these datasets to the testing alternatives of the CIFAR datamodule
• Add the reference to the documentation
• Add torch-uncertainty in PapersWithCode's dataloaders

See https://github.com/ENSTA-U2IS-AI/torch-uncertainty/actions/runs/8408560029/job/23024874240

Remove randomness.

To close this issue, we should at least:

• Check that the regression routine is behaving correctly
• Add NYUv2
• Add KITTI
• Add MUAD
• Add simple baselines for depth regression

Paper: https://arxiv.org/abs/2108.00968

Implementation: https://github.com/giannifranchi/deeplabv3-superpixelmix

More particularly, the transform is here: https://github.com/giannifranchi/deeplabv3-superpixelmix/blob/master/datasets/cityscapes_mix.py#L13

Original code: https://github.com/giannifranchi/LP_BNN

Original paper: https://arxiv.org/abs/2012.02818

In one of my latest training, I had the following error in plotting_utils.py, line 95:

val_oh = torch.nn.functional.one_hot(val.long(), num_classes=10)
RuntimeError: Class values must be non-negative.

I don't have much more information about this, but we should check if some cases lead to such an error.

TinyImageNet-C can be downloaded here: https://berkeley.app.box.com/s/6zt1qzwm34hgdzcvi45svsb10zspop8a

• Check that it can be downloaded automatically
• Create the corresponding dataset
• Add the reference to the documentation

A single test (over local tests and GitHub actions) has failed due to the Bayesian layers. Let's look into this.

Source:
https://github.com/ENSTA-U2IS/torch-uncertainty/actions/runs/5358435126/attempts/1

See https://torch-uncertainty.github.io/auto_tutorials/tutorial_scaler.html#sphx-glr-auto-tutorials-tutorial-scaler-py.

This goes for all tutorials.

Now that we support regression, it could be interesting to implement simple methods, such as Deep Evidential Regression (DER).

This method would be relatively straightforward to implement. Here is a possible roadmap:

• Add Deep Evidential Regression's loss
• Change "dist_estimation" parameter of the regression routine by an integer parameter to choose the number of parameters that we have to predict for the distribution
• Add tests
• Add a tutorial to explain how to use DER
• Add the reference to the documentation

The token is not valid, and we cannot upload the coverage cf. this GitHub action.

There is no possibility of enforcing the installation of torch with cpu only for now.
Implement this feature.

Some tests that fail locally pass in the CI/CD. They correspond to the pull request target synchronization.

Currently, validation may fail even for relatively small models such as DeepLabV3 when using large image crops. Calibration Error is the worst metric but others seem to also have huge impacts on the VRAM.

Hello, after studying your paper, it has been very inspiring for my work. I still have some questions that I would like to consult with you

For the learning rate issue of training, what your code means is that the loss function is the average value of the losses of all experts. If there are four experts, then for each expert, the actual loss is divided by four, which means that when backpropagation is used to calculate the gradient, it will also be divided by four. Do you need to initially set a learning rate that is four times larger than the single model

Hello, thank you for your open source code to my research has a great inspiration and help. One small question I have is that PackedLinear and PackedConv2d in packed_layers.py don't seem to correspond to the latest github repository and the folders associated with them don't seem to be updated after pip installation. For example, the PackedLinear function in the source code of pip installation has no alpha parameter. Could you please update it to facilitate the use of pip installation？

Hi, thanks for your excellent work and sharing codes. I have a small question. If I apply packed ensemble, the pre-trained parameters of the ResNet on the ImageNet can't be used. Is it right?

I'm running following command on cifar10

python3 resnet.py --version mimo --arch 18 --accelerator gpu --device 1 --benchmark True --max_epochs 75 --precision 16 --root "./data/" --num_estimators 4

but it raises error RuntimeError: The size of tensor a (512) must match the size of tensor b (128) at non-singleton dimension 0. I dive into the code and find that it may be related to this line, should we remove it? @o-laurent Could you help with this?

Edit by maintainer:

• Either add an "mc_dropout" parameter to the CLI or to the models (we could have dropout during training + MIMO or else, especially for wideresnets)

Display the distance with the optimal calibration bins.

For instance, with dashed red histograms.