The scripts presented here are part of my Master thesis, the results of which were published in the paper A Deep Learning Wireless Transceiver with Fully Learned Modulation and Synchronization. They allow training and evaluation of autoencoders, that are able to transmit data over a wireless channel, using TensorFlow (probably 1.4, 1.5 or 1.6) and the in 2019 still standalone Keras frameworks. The trained autoencoder can be imported into GNURadio for over-the-air transmission using Tensorflow GNURadio Blocks.
The export files of four examplary trained autoencoder used to be found here. Their names indicate the ratio of transmitted bits per complex baseband samples, for example AE-7/8 transmits 7 bits using 8 samples. Our experiments showed that AE-7/16 and AE-8/8 perform best, both over the air and when evaluated over the channel model they were trained on.