lazy load training

[paulivii]

Hi all, as me and my team work with arm devices, we have a special focus on speed and memory consumption. More precise, we want to perform model trainings on an edge device. That is currently possible but requires relatively large memory. E.g. a classifier yolo training in our configuration requires ~1.5Gb and a detector training ~2.2Gb of memory. However, a not neglectable part of that comes from imports that are actually not necessarily required such as Tensorflow, Keras, HuggingFace, OpenVINO. Therefore we are wondering if it is possible to only "lazy-load" those requirements so that they are not imported in every training?

General benefits for the community:

• memory reduction
• faster startup times of python processes (importing takes time)

Thanks in advance

[glenn-jocher]

@paulivii thank you for your suggestion. Currently, the Ultralytics framework doesn't support lazy loading for these dependencies. However, optimizing imports for specific use cases is a valuable idea, and we appreciate your feedback. For further insights, you might explore customizing your environment to exclude unnecessary packages.