rstojnic
rstojnic
In some cases the first batch might be too small, or simply contain no interesting predictions (eg empty or all zeros). Run hash calculation should catch this, produce a warning...
Use PEP-8 standard for formatting code, perhaps run the whole code through a formatter or install the pre-commit hooks ?
We should add latest pytorch/torchvision installation from github into requirements, then eliminate all copy-pasted code.
TODO sotabench lib: - remove benchmark() function from `benchmark.py` - move deps to requirements - evaluation.json should be made if some ENV variable is set, otherwise pprint something - for...