Config files refactor, Examples polish, Perturbation in GraphDefinition #603
Conversation
…cation warning. To not break backwards compatibility.
fixed typo
|
Since I've also contributed to this PR maybe it would be better if @Aske-Rosted also had a look at it? |
examples/04_training/README.md
Outdated
| --model-config configs/models/example_direction_reconstruction_model.yml \ | ||
| --prediction-names zenith_pred zenith_kappa_pred azimuth_pred azimuth_kappa_pred | ||
| ``` | ||
|
|
||
| **`02_train_model_without_configs.py`** Shows how to train a GNN on neutrino telescope data **without configuration files,** i.e., by programatically constructing the dataset and model used. This is good for debugging and experimenting with different dataset settings and model configurations, as it is easier to build the model using the API than by writing configuration files from scratch. For instance, try running: | ||
| **`02_train_model_dynedge_from_config.py** Shows how to train a GNN on neutrino telescope data **using configuration files** to construct the dataset that loads the data and the model that is trained. This is the recommended way to configure standard dataset and models, as it is easier to ready and share than doing so in pure code. This example can be run using a few different models targeting different physics use cases. For instance, you can try running: |
There was a problem hiding this comment.
I think you mean 03_...
|
@Aske-Rosted Could you give this a look at some point? As @AMHermansen both him and I contributed to this, so a 3rd pair of eyes would be brilliant. |
There was a problem hiding this comment.
if the Not version proof indicates that
model.save_state_dict()
model.save_config()
is more robust than
model.save()
and that this might fail due to a version conflict then I would consider changing the order of the calls
There was a problem hiding this comment.
Thanks for this. What the comment means is not that executing these lines might lead to an error, but that archiving your model as .pth file under one version of graphnet can give you problems if you try to load it in a different version of graphnet. I'll update the comment to reflect this better.
…s_perturbed_dataset_refactor Config files refactor, Examples polish, Perturbation in GraphDefinition
A continuation of #591.
Closes #583 and #588