The librosa.effects module provides wrappers to hpss that can give you the time domain signal instead of the spectrogram.
Hpss will do an okay job of separating drums, depending on the mix, but you might also want to look at pre trained source separation models like spleeter or open-unmix.