I’m doing reinforcement learning on Atari games with Tensorflow and Gymnasium, and realised that it takes more iterations to converge when using tf.function with lazy execution compared to eager execution. Though lazy execution are still faster overall, even with the additional iterations.