heth27

Checkpoint every_n_steps reruns epoch on restore

3

### Bug description The checkpoint callback is run before ```batch_progress.increment_completed()``` in training_epoch_loop's ```advance``` method. Thus in the checkpoint checkpoint['loops']['fit_loop']['epoch_loop.batch_progress']['total']['completed'] e.g. 9 is one smaller than for example checkpoint['loops']['fit_loop']['epoch_loop.batch_progress']['total']['processed'] e.g. 10...

Smoothing in tqdm progress bar has no effect

2

### Bug description The option smoothing when creating progress bars in TQDMProgressBar has no effect in the default implementation, as _update_n only calls bar.refresh() and not the update method of...

### Bug description training mode does not get switched on automatically in the training_step, so if the model is put into eval mode accidentally by the user it stays this...

### Bug description Loading a model from a checkpoint `model = TestLitModel.load_from_checkpoint(ckpt_path)` changes the pytorch random state as I guess the model is initialized and then the weights are restored....

### Description & Motivation Knowing the norm of weights was necessary to debug float16 training for me. ### Pitch from lightning.pytorch.utilities import grad_norm norms = grad_norm(self.layer, norm_type=2) something like this...

procrustes alignment for pytorch

3

### 🚀 The feature Orthogonal procrustes alignment ### Motivation, pitch Procrustes alignment is a staple when calculating metrics for 3d human pose estimation, but there seems to be no library...

## 📚 Documentation What is the difference between kornia and pytorch3d, https://pytorch3d.readthedocs.io/en/latest/index.html To me they seem to be quite similar in what they try to achieve, so a hint in...