I replaced the torchmetrics class AverageMeter with MeanMetric, which is essentially the same.
I added the variable full_state_update = False at every PyCave's custom metric as suggested here. I set all of them to False as it doesn't seem to influence the results of the tests.
Within KmeansPlusPlusInitLightningModule.nonparametric_training_epoch_end it was giving me errors because sometimes the choice variable is an empty tensor, and indexing a zero-dimensional tensor is not allowed since pytorch>=0.5. Is it supposed to be empty? Is there something breaking that is not captured by the tests? I added a simple if there, but if you have a more elegant solution let me know.
I run the test on both pytorch-lightning=1.6.0, torchmetrics=0.6.0 (minimum required) and pytorch_lightning=1.8.1, torchmetrics=0.10.2 (the most updated so far) and everything seem to work well.
First of all, amazing project!
AverageMeterwithMeanMetric, which is essentially the same.full_state_update = Falseat every PyCave's custom metric as suggested here. I set all of them toFalseas it doesn't seem to influence the results of the tests.KmeansPlusPlusInitLightningModule.nonparametric_training_epoch_endit was giving me errors because sometimes thechoicevariable is an empty tensor, and indexing a zero-dimensional tensor is not allowed since pytorch>=0.5. Is it supposed to be empty? Is there something breaking that is not captured by the tests? I added a simple if there, but if you have a more elegant solution let me know.I run the test on both pytorch-lightning=1.6.0, torchmetrics=0.6.0 (minimum required) and pytorch_lightning=1.8.1, torchmetrics=0.10.2 (the most updated so far) and everything seem to work well.
Have a nice day!