1e15eeebbe3f9647f4a024ce9d44f5f6bb37b179,tsne/tsne.py,,gradient_descent,#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,612

Before Change

            // Correct the KL divergence w.r.t. the exaggeration
            error = (error - np.sum(P) * np.log(exaggeration)) / exaggeration

            print("Iteration % 4d, error %.4f" % (iteration + 1, error))

        grad_direction_flipped = np.sign(update) != np.sign(gradient)
        grad_direction_same = np.invert(grad_direction_flipped)
        gains[grad_direction_flipped] += 0.2

After Change

        if should_call_callback:
            // Continue only if all the callbacks say so
            should_continue = all((bool(c(iteration + 1, error, embedding)) for c in callbacks))
            if not should_continue:
                // Make sure to un-exaggerate P so it"s not corrupted in future runs
                if exaggeration != 1:
                    P /= exaggeration

Instances