You are developing a machine learning model intended to classify whether X-ray images indicate bone fracture risk. The model employs a ResNet architecture and has been trained on Google's Vertex AI platform, utilizing a TPU as an accelerator. Despite this, you are unsatisfied with the training time and memory usage. You need to iterate on your training code quickly but prefer to make minimal changes to the existing codebase. Additionally, you aim to minimize any potential impact on the model’s final accuracy. Which of the following actions should you take?