Implement the Cross-Entropy loss for multi-class classification.
Given class probabilities (after softmax) and integer class labels:
$$\text{CE} = -\frac{1}{N} \sum_{i=1}^{N} \log(p_{i, y_i})$$
where $p_{i, y_i}$ is the predicted probability for the true class of sample $i$.
Input:
probs: a 2D tensor of shape (N, C) with predicted probabilities (rows sum to 1) targets: a 1D integer tensor of shape (N,) with class indices Output: A scalar representing the mean cross-entropy loss
Note: Add a small epsilon (1e-7) to avoid log(0).