TabPFN
TabPFN copied to clipboard
Start benchmarking
Started this to gather some feedback: Metrics to gather from benchmarking:
-
Speed
-
Accuracy
-
Speed: Two types of measures
- End to end: It is fine to use
timefunction as long as we are measuring process that take at least around 1 sec (otherwise it might become too unreliable to read). - Finer Measure: This can be done with ``cProfile` package from python that give stats details at the level function calls Speed is also machine dependent (Cpu/Gpu specs, memory). So, it needs these specs In speed measurement, we can also measure speed as we scale different parameters (num_sample, num_features, num_classes)
- Accuracy: Need to establish some datasets to measure these on.
Another metric: Memory requirements ( but that seems somewhat managed by pretraint constraints)
Dear @mert-kurttutan apologies we didn't have the capacity to work with you on the PR. Thank you for the thoughtful work back then 🙏 We will keep the PR in mind when looking more into the workflow.