We introduce DeepNovoV2, the state-of-the-art neural networks based model for de novo peptide sequencing. Contrary to existing models like DeepNovo or DeepMatch which represents each spectrum as a long sparse vector, in DeepNovoV2, we propose to directly represent a spectrum as a set of (m/z, intensity) pairs. Then we use an order invariant network structure (T-Net) to extract features from the spectrum. By representing spectrums as sets of peaks, we argue that our method is more straightforward and do not have the accuracy-speed/memory trade-off problem. Our experiments show that comparing to the original DeepNovo model, DeepNovoV2 has at least 15% relative improvement on peptide accuracy.