Learning with kernels is an often resorted tool in modern machine learning. Standard approaches for this type of learning use a predefined kernel that requires careful selection of hyperparameters. To mitigate this burden, we propose in this paper a framework to construct and learn a data-dependent kernel based on random features and implicit spectral distributions (Fourier transform of the kernel) parameterized by deep neural networks. We call the constructed network {\em KernelNet}, and apply it for deep generative modeling in various scenarios, including variants of the MMD-GAN and an implicit Variational Autoencoder (VAE), the two popular learning paradigms in deep generative models. Extensive experiments show the advantages of the proposed KernelNet, consistently achieving better performance compared to related methods.