The analysis of the waveforms of photo-multiplier tube (PMT) is essential for high precision measurement of position and energy of particles' interaction in liquid scintillator (LS) detectors. JUNO is a next-generation large volume liquid scintillator neutrino experiment with a designed energy resolution of 3% @1 MeV. The accuracy of the reconstruction of the number of photoelectron (nPE) is one important key of achieving the best energy resolution. This work introduces the machine learning-based nPE estimation methods, including supervised learning depended on electronic simulation and data-driven weakly supervised learning. The calibration parameters of LS and PMT responses are used to generate training waveforms for supervised learning. The photon counting performances of different methods will be presented.