评估标准
R = T / RT P = T / PE T = equal(PT, RT), PT
For brevity, let x = logits, z = labels. The logistic loss is
z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x)) = z * -log(1 / (1 + exp(-x))) + (1 - z) * -log(exp(-x) / (1 + exp(-x))) = z * log(1 + exp(-x)) + (1 - z) * (-log(exp(-x)) + log(1 + exp(-x))) = z * log(1 + exp(-x)) + (1 - z) * (x + log(1 + exp(-x)) = (1 - z) * x + log(1 + exp(-x)) = x - x * z + log(1 + exp(-x))
For x < 0, to avoid overflow in exp(-x), we reformulate the above
x - x * z + log(1 + exp(-x)) = log(exp(x)) - x * z + log(1 + exp(-x)) = - x * z + log(1 + exp(x)
Hence, to ensure stability and avoid overflow, the implementation uses this equivalent formulation
max(x, 0) - x * z + log(1 + exp(-abs(x)))