Open n17s opened 7 years ago
Was this ever resolved? I am having the same issue in 2.6. The model executes via a Trainer
object during a training routine. But on trying to evaluate the forward pass for validation, BatchNorm
produces NaN
values. After running with debug_model
I have the same error as above.
Continuing on the debug_model issues (c.f. #2232 and #2231) debugging a model with batchnorm doesn't work. The repro is: