Recent developments in applying deep learning techniques to train end-to-end communication systems have shown great promise in improving the overall performance of the system. However, most of the current methods for applying deep learning to train physical-layer characteristics assume the availability of the explicit channel model. Training a neural network requires the availability of the functional form all the layers in the network to calculate gradients for optimization. The unavailability of gradients in a physical channel forced previous works to adopt simulation-based strategies to train the network and then fine tune only the receiver part with the actual channel. In this letter, we present a practical method to train an end-to-end communication system without relying on explicit channel models. By utilizing stochastic perturbation techniques, we show that the proposed method can train a deep learning-based communication system in real channel without any assumption on channel models. © 1997-2012 IEEE.