Adding final version of entropy entropy acquisitions with test cases

[rishabh-mondal]

Implemented Entropy acquisition and test cases on commit 2944d64

Files added:

1. astra/torch/al/acquisitions/entropy.py: contain implementation of entropy acquisition
2. astra/torch/al/strategies/deterministic.py: modified this file as per the need for test case purpose
3. astra/torch/al/strategies/ensemble.py: modified this file as per the need for test case purpose
4. astra/torch/al/strategies/mc.py: modified this file as per need for test case purpose

Passes all test cases, including those already existing 2944d64
Explanations:

class EntropyAcquisition(MCAcquisition, EnsembleAcquisition, DeterministicAcquisition):
This line defines a new Python class called EntropyAcquisition. This class inherits from three parent classes: MCAcquisition, EnsembleAcquisition, and DeterministicAcquisition.

def acquire_scores(self, logits: torch.Tensor): if logits.dim() == 3:
If the logits tensor has three dimensions, this line computes the log probabilities (log-softmax) along the last dimension (dim=2) using PyTorch's F.log_softmax function and stores the result in the log_probs variable.

entropy = -torch.sum(torch.exp(log_probs) * log_probs, dim=2)
This line calculates the entropy of the logits.
entropy = torch.sum(entropy, dim=0)
This line sums up the calculated entropies across all samples in the sequence by summing along the first dimension (dim=0), resulting in a single entropy list.

else: log_probs = F.log_softmax(logits, dim=1) entropy = -torch.sum(torch.exp(log_probs) * log_probs, dim=1) return entropy
If it is deterministic, it is just returns a list of entropy.