Hey team, quick question about reproducibility of ...
# squads
m
Hey team, quick question about reproducibility of TimeGPT. When using fine-tuning, are small differences with the same output to be expected?
If yes, can you explain why and in which order of magnitude they're expected to be? (e.g. 1e-3)
y
Oh, yes. I think Ezdrasz encountered a similar issue. This is Jose’s explain:
Copy code
We have different GPU models so depending on which one runs the process the results can vary, each one may use a different algorithm for matmul so you end up with slightly different results in each forward pass and 700 finetune steps can cause bigger differences.
we have a relative tolerance of 0.1% and absolute tolerance of 0.0001
m
thank you! this is exactly what I needed.
🙌 1