Added wandb, metrics, mask2former

[Alighieri1231]

So we go again, this is my first pull request so sorry if I need to be more detailed; I really like the project!

• Added Dice score as metric
• Added IoU and Dice score for foreground and classes
• Added IoU calculation from nnUnet library

The reason why I added so many dice and IoU is because all of them are different! So I need to analyze which one is the correct but in the meantime I wanted to train all the cases (you can see the average and std of the curves because I did cross-validation) .

• Added wandb logs for improvement of visualization

• Added mask2former as a new decoder:

• For this I needed to add dinov3 library, in order to run the code you need to compile 'MultiScaleDeformableAttention' once, for doing this you can do:

%cd /dinov3-finetune/dinov3/eval/segmentation/models/utils/ops/ !python setup.py build install

Once you do that you will be able to run mask2former. An example of running this new code is in run_gpu1_mask2former.sh

I achieve similar results as using LoRA with FPN and the head, I did not have time to train the mask2former with LoRa, my dataset was not too big so maybe saturation is gonna happen.

• Added 32-bit precision for increased batch size.
• Added some sh files to run different experiments at a time:

I am including two SLURM files if you want to run several experiments in a cluster at a same time (sh files).

Let me know If I need to be more organized, I would love to have this project growing! Thanks. What I am doing for the next iteration is implementing classic U-nets as benchmarks, so we can have everything in one repository.

[RobvanGastel]

Hi again, thank you for adding all these components, I am glad it was interesting for you! Maybe we or I could help you here, could clean up the code and some of the slurm bash scripts as different users might have different ways of running their experiments.