diff --git a/docs/src/architectures/pet.rst b/docs/src/architectures/pet.rst index c60393609..8b8216ce7 100644 --- a/docs/src/architectures/pet.rst +++ b/docs/src/architectures/pet.rst @@ -6,8 +6,8 @@ PET .. warning:: The metatrain interface to PET is **experimental**. You should not use it for - anything important. You can also fit PET from `here - `_. + anything important. You can also fit PET using native scripts (not experimental) + from `here `_. Installation @@ -121,6 +121,17 @@ suffice. However, for particularly large datasets, increasing ``step_size`` may necessary to ensure complete convergence. The hyperparameter controlling the ``step_size`` of the StepLR learning rate scheduler is called ``SCHEDULER_STEP_SIZE``. +It is worth noting that the default ``step_size`` is quite large. Thus, it is normal +if, when fitting on V100, which is quite slow, there is no event of lr rate decrease +during the first day or even during a couple of days. In addition, for some datasets, +the fitting might take longer than for others (related to inhomogeneous densities), +which can further postpone the first event of lr decrease. + +The discussed convergence above, especially in terms of the total duration of fitting, +should preferably be checked on log-log plots showing how the validation error depends +on the epoch number. The raw log values are typically hard to extract useful insights +from. + For hyperparameters like ``SCHEDULER_STEP_SIZE``, ``EPOCH_NUM``, ``BATCH_SIZE``, and ``EPOCHS_WARMUP``, either normal or atomic versions can be specified. ``SCHEDULER_STEP_SIZE`` was discussed above; ``EPOCH_NUM`` represents the total number