Create acquisition functions

[patel-zeel]

Refer to this paper for most functions below.

• Create a folder structure for acquisition functions - @patel-zeel
• Add asserts to ensure only supported pairs of acquisitions and strategies work together - @patel-zeel
• Entropy - @rishabh-mondal
• BALD - @aditiagarwal-02
• Max. variance ratios - @Khush24Shah
• Mean STD - @VannshJani
• Learning to Loss - @Madhav-Kanda
• Diversity (for loop)- @jaiswalsuraj487
• BatchBALD - @patel-zeel

Create a separate file for each of these strategies in astra/torch/al/acquisitions folder. See for example, https://github.com/sustainability-lab/ASTRA/blob/main/astra/torch/al/acquisitions/random.py

Your acquisition function should inherit from all base classes it supports. For example, Entropy class should inherit from MCAcquisition and EnsembleAcquisition:

from astra.torch.al.acquisitions.base import MCAcquisition, EnsembleAcquisition

def EntropyAcquisition(MCAcquisition, EnsembleAcquisition):
    ...

All of you should

• Create the acquisition implementation files.
• Add appropriate tests for them in astra/tests directory in relevant folders.
• Add a small notebook showing a minimum working active learning example using your acquisition.

[patel-zeel]

I have added DeterministicStrategy now. Acquisitions such as Entropy should check if the input is 3d (mc_dim, pool_dim, n_classes) and take mean over the first dimension. If input is 2d (pool_dim, n_classes), simply compute the acquisition and return the scores.

[aditiagarwal-02]

Since MC BALD and BALD have been assigned differently, BALD will only inherit from DeterministicStratergy? Basically, should we have two separate files for BALD and MC BALD?

[patel-zeel]

@aditiagarwal-02 No, you can inherit from all relevant classes at once. I have updated the assignment accordingly.