Closed DavidDiazGuerra closed 10 months ago
Hi again,
The problem disappeared after downgrading Lightning to 1.9.4. I guess there are some changes in the 2.0 version that breaks the X-UMX training routine of Asteroid.
Best, David
Hello 👋
I had the same problem, it seems that the function definition is def lr_scheduler_step(self, scheduler, metric)
in version 2.0 instead of def lr_scheduler_step(self, scheduler, optimizer_idx, metric)
in previous versions (see the custom learning rate schedulers documentation here).
@mpariente Should we change this and update the requirements to pytorch-lightning>=2.0.0
?
I'm not sure this code has an easy backward compatible fix. So we should
pytorch-lightning>=2.0.0
with asteroid 0.7.0.I'm closing this issue since it has been fixed with #682 and the release 0.7.0.
Hello,
I'm trying to use the receipt for training the X-UMX model on the musdb18 dataset but the following exception raises after finishing the validation of the first epoch:
I am using Pytorch 2.0.0 and Lightning 2.0.2. In order to make the model work with Pytorch 2.0.0 I've made some minor changes in the X-UMX model (see #662) and I've reduced the batch size to make the model fit in my local GPU, but I don't think I've made any changes that should affect the Lightning training.
Maybe this could be related to be using recent versions of Pytorch and Lightning? I have no experience working with Lightning.
Best regards, David