✨ Add Monte-Carlo Dropout

[badrmarani]

This PR fixes #40. The main changes are:

• Create a new baseline.
• Find a solution to apply dropout in eval mode.
• Add tests.
• Add a tutorial to explain how to use the baseline.
• Add the reference to the documentation.
• Add the parameter enable_last_layer_dropout to enable last-layer dropout.
• Extend this method to VGG and WideResNet.

[o-laurent]

Hi! Thanks for the great work, again!

I'll have a more thorough look at it next week. I've skimmed your PR and am just wondering how we could adapt your baseline to the different variants of MC Dropout. It may be a question worth considering, given your interest in its variants!

Then, there is also the question of the number of inference passes. I agree that it is more optimal to avoid the for loop and increase the batch size, but it may lead, in some cases, to increased VRAM needs (for instance, with 1000 passes?). We could let the user select the test batch size if wanted (that could be reduced to 1, for instance, if we want to compute the mean on many passes). It seems to be the simplest solution, but it is not completely safe & it adds a new parameter to the CLI.

[o-laurent]

Thanks for the PR. I will make some modifications, but having it in the library is great. If you want, you can join the Discord server whose link I added in the readme of the dev branch!

[o-laurent]

I had to make a few changes to avoid errors. Merging now.