You are training a ResNet model on Google AI Platform using TPUs to visually categorize types of defects in automobile engines. During the training process, you use the Cloud TPU profiler plugin and observe that the process is highly input-bound, meaning that the training is limited by the speed at which data is fed into the model. You want to address this bottleneck to speed up your model's training process. Which modifications should you make to the tf.data dataset?