Closed abs-xyz closed 1 year ago
Perhaps torchinfo can catch the exception and suggest a fix using device
. I'm not very interested in cloning the model instance to support this mistake, since it seems like it is easily avoided by reading the documentation.
Is your feature request related to a problem? Please describe. I was facing the issue described here. I wanted to train on CPU, although CUDA was available. It took me a long time to figure out that the problem was due to my incorrect usage of
summary
.Describe the solution you'd like It would be great if the forward pass won't lead to the same model weights being put on the GPU. Maybe the model instance can be cloned and then this forward pass be done? This would not disrupt the further training, even if someone forgets to set the
device
argument.