computervision-recipes
computervision-recipes copied to clipboard
microsoft
[BUG] check why linux tests now same speed as windows, and why/if linux test got slowr
Description
| Dataset name | Time env installation | Time tests | #tests | Time integration test for IC test_03_notebook_run | Time unit test for IC test_03_notebook_run | Comment | ||
|---|---|---|---|---|---|---|---|---|
| Linux GPU | 3:20min | 22min | 74 | 183sec | 19 sec | On Aug 22nd, the IC integration 03 test took 57 sec | ||
| Linux CPU | 4:30min | 5min | 66 | 27 sec | ||||
| Windows GPU | 9:20min | 10min | 66 | 16 sec | ||||
| Windows CPU | 14min | 11min | 66 | 26 sec | ||||
| AzureML notebooks |
@jiata Can I close this? Not sure if it's related, but I made some updates which speed up the windows tests (mainly for image classification and image similarity, but also for the other scenarios. Turns out PyTorch's dataloader when using >1 workers in windows gets very slow. Hence the number of workers is now automatically set to 1 for windows machines (and if on Binder, that was to fix an otherwise occurring error)
