-
In one project, the transformers package has to be version:
transformers==3.5.1, and Pytorch has be >= 1.8.0 - Installing the legacy version may have some conflicts, but anaconda should solve the conflicts -
Using later pytorch would give:
ImportError: cannot import name 'SAVE_STATE_WARNING' from 'torch.optim.lr_scheduler'as this is removed in pytorch recently, but the older transformers package calls it. -
The torch module does check if a lower version of Pytorch is installed to avoid conflicts, but higher version is not checked by the module. i.e:
1
2
3
4
if version.parse(torch.__version__) <= version.parse("1.4.1"):
SAVE_STATE_WARNING = ""
else:
from torch.optim.lr_scheduler import SAVE_STATE_WARNING
- Fix by checking if higher version of pytorch is used in
transformers.trainer_pt_utils.py:
1
2
3
4
5
if version.parse(torch.__version__) <= version.parse("1.4.1") or version.parse(torch.__version__) >= version.parse("1.8.0") :
SAVE_STATE_WARNING = ""
else:
from torch.optim.lr_scheduler import SAVE_STATE_WARNING
