We propose and study a kernel estimator of a density in which the kernel is adapted to the data but not fixed. The smoothing procedure is followed by a location-scale transformation to reduce bias and variance. The new method naturally leads to an adaptive choice of the smoothing parameters...