Bug in Cross-Entropy Implementation inside `Scholar.Linear.LogisticRegression`?

[krstopro]

I believe there is a reduction missing (probably Nx.mean/2) in the implementation of cross-entropy inside Scholar.Linear.LogisticRegression.

scholar/lib/scholar/linear/logistic_regression.ex

Line 188 in 975938a

-Nx.sum(ys * log_softmax(Nx.dot(xs, coeff) + bias), axes: [-1])

xs is a tensor of shape {num_samples, num_features}, ys is a tensor of shape {num_samples, num_classes} (one-hot encoded). Hence, the expression above evaluates to a tensor of shape {num_samples}. I don't know how exactly does Polaris handle vectors, but I guess the loss should be a scalar.