search

AI4HealthUOL / SSSD

Repository for the paper: 'Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models'
MIT License
270 stars 47 forks source link

Training times/epoch stats #11

Closed nick-torenvliet closed 1 year ago

nick-torenvliet commented 1 year ago

Random question about your paper. In appendix b.3 you quote some training times. I am wondering about the hardware. Can you provide RAM, CPU(# and type), GPU (# and type).

I am wondering if you could discuss training times, and hardware required, to produce the various SOA results you have quoted throughout the rest of the paper.

juanlopezcode commented 1 year ago

Hi ram: 8 gb cpu: 8 cores intel(r) xeon(r) platinum 8362 cpu @ 2.80ghz gpu: a30 24gb and ~3800 cuda cores

gminorcoles commented 1 year ago

I have not much experience with PyTorch and when I try to run train.py on an A100 with 40gb GPU ram, using the electricity dataset, it fails saying there is not enough GPU memory available.

juanlopezcode commented 1 year ago

Hello @gminorcoles, I would go with smaller batch size, but the model fits in a30s cards so it should fit in yours, unfortunatelly, debugging more than that would go beyond my assitance capabilities