leiterenato / tf-guides Goto Github PK

Thanks for writing the guide! I managed to find it from https://www.kaggle.com/code/leiterenato/high-performance-pipelines-with-tf-data-and-gcp

This might not be the best place to seek advice, however i think there are not alot of tutorial for training on google cloud platform. I also referred to this doc

Currently, I'm working on a research project which utilized GCP cloud bucket as datastorage, datas are loaded in the form of rlds, size of approx 100mb each. The training is conducted on TPU vm4. With some optimizations by manually tuning the parralel_calls, parralel_reads, buffer size etc. on each operation, there's still a bottleneck on the dataloader. Currently the half of the time is dataloading and half is training. It is obvious that more optimization can be done.

I used profiler to check on each dataload operation, and it seems reasonable. Also, I notice that the it/s and CPU usage fluctuates abit, (i have reached almost max CPU ram usage). Since there are so many knobs that can be tuned, do you have any advice to improve this dataloading operation?

Would very much appreciate it if you can give me some quick tips! Thanks 🙏