Pulse shaping for coherent optical fiber communication has been an active area of research for the past decade. Most of the early schemes are based on classic Nyquist pulse shaping that was originally intended for linear channels. The best known classic scheme, the split digital back-propagation (DBP), uses joint pre-distortion and post equalization and hence, a nonlinear transmitter (TX); it, however, suffers from spectral broadening on the fiber due to the Kerr-effect. With the advent of deep learning in communications, it has been realized that an Autoencoder can learn to communicate efficiently over the optical fiber channel, jointly optimizing geometric constellations and pulse shaping - while also taking into account linear and nonlinear impairments such as chromatic dispersion and Kerr-nonlinearity. E.g., arXiv:2006.15027 shows how an Autoencoder can learn to mitigate spectral broadening due to the Kerr-effect using a trainable linear TX. In this paper, we extend this linear architectural template to a scalable nonlinear pulse shaping consisting of a Convolutional Neural Network at both transmitter and receiver. By introducing a novel $\gamma$-lifting training procedure tailored to the nonlinear optical fiber channel, we achieve stable Autoencoder convergence to pulse shapes reaching information rates outperforming the classic split DBP reference at high input powers.