pytorch - 💡(How to fix) Fix Deterministic implementation of CTC Loss [2 comments, 3 participants]

Fix Plan

To fix the RuntimeError caused by the non-deterministic implementation of CTC Loss, you can use the warn_only=True option when setting deterministic algorithms. Here are the steps:

• Set torch.use_deterministic_algorithms(True) with warn_only=True:

import torch
torch.use_deterministic_algorithms(True, warn_only=True)

Alternatively, you can turn off determinism just for the CTC Loss operation:

• Use torch.backends.cudnn.deterministic = False before the CTC Loss operation:

import torch
torch.backends.cudnn.deterministic = False
# Your CTC Loss operation here

• Remember to set torch.backends.cudnn.deterministic = True after the operation if needed.

Verification

To verify that the fix worked, run your code with the modified deterministic algorithm settings and check that the RuntimeError is resolved.

Extra Tips

Note that using non-deterministic algorithms may affect the reproducibility of your results. If determinism is crucial for your application, consider filing an issue at https://github.com/pytorch/pytorch/issues to request deterministic support for the CTC Loss operation.