I'd love to ditch Tensorflow and just use PyTorch + Tensorboard. Could always support both as backends.