Sagemaker downloads huggingface model image every time on running huggingface_estimator.fit()
In a local notebook, you could always save the model locally and refer to it using a file path (e.g. /Users/sanjanag/models/bert-base). But since this is a job running on ECS, I don’t think you get that choice. It runs from scratch every time.
Hey @sanjanag,
Thank you for opening the thread. Could describe a bit more was your question is?
Model downloading when starting the training takes a few seconds since you are loading with 50-100MB/s.
Or can’t you load the model on training start due to compliance guidelines?