automl
Print warning for larger datasets when run on CPU
Describe the workflow you want to enable
Creating this new issue to be able to close #100
Describe your proposed solution
Print warning for larger datasets when run on CPU (>1000 samples? or rather when the estimated runtime is very long? > 10 min). Warning should say that user can try our API client instead or move to GPU.
Describe alternatives you've considered, if relevant
No response
Additional context
No response
Impact
None
I think we add below code to regressor.py in fit method in order to warn users on CPU when the dataset is large
import warnings
if self.device == 'cpu' and X.shape[0] > 10000:
warnings.warn("Running on CPU with a large dataset may be slow. Consider using a GPU.")
@noahho @LeoGrin please tell me whether I am thinking in right direction.
Yes exactly, I'd just put this at 1000 samples and also include a reference to using the API if GPU not available https://github.com/PriorLabs/tabpfn-client
Longer term, I'd make a larger change where we estimate the time taken, similar to: https://github.com/PriorLabs/TabPFN/blob/main/src/tabpfn/model/memory.py#L107
@noahho I had raised an PR applying "Yes exactly, I'd just put this at 1000 samples and also include a reference to using the API if GPU not available https://github.com/PriorLabs/tabpfn-client" changes, please review it. I will try to implement long term solution also.
Hi, thanks for opening this issue! I'm interested in working on this feature. I have some experience with Python and optimizing ML workflows, so I’m excited to contribute.
Before I get started, I have a couple of questions to ensure I fully understand the requirements:
For the warning threshold, would you prefer to trigger it based on a fixed sample count (e.g., >1000 samples) or an estimated runtime (e.g., >10 minutes), or perhaps a combination of both? Should the warning be printed as a terminal output (e.g., using Python’s logging module) or integrated into the API’s response?
Looking forward to your guidance so I can implement this in a way that best fits the project's workflow. Thanks!
Best regards, Rishi
