Is this faster even after applying the optimizations that reduce VRAM usage? (some of which the Keras version seem to lack)