Experiments have been re-organized per scenarios :
“Ecologists”
In this scenario, users possess a dataset of their own on which they want to train a model while being able to easily customize the training process and the model architecture.
“Inference”
In this scenario, users possess both a test dataset and a trained model. They want to simply test the performances of their model with relevant metrics.
“Kaggle”
In this scenario, users are participants of a GeoLifeCLEF challenge and look for plug-and-play examples of data loaders and/or training examples, specifically developed for the challenge dataset.
Updated config files with new paths and cleaned example files. All example files now use Hydra v1.3 which does not change the current working directory to hydra.run.dir, so all paths given to config files should be relative to the experiment directory (e.g.: malpolon/examples/ecologists/sentinel-2a/).
Malpolon's engine classes (in particular the datasets classes) have been split for better encapsulation.
Metrics calling have been updated. A new dictionary was added in malpolon.data.utils containing torchmetrics names as keys and functional metrics callable as values. Hence users only need to add an entry in their config file with the same name as the dict keys + the torchmetrics kwargs under it. If a user wants to use a torchmetrics functional metric not listed in the dictionary, they can choose a custom name as entry for the config file and need to a sub-entry callable: <Fmetrics.X.X>.
[x] Inference
Added inference computing methods and methods to output inference predictions (metrics, logs, tensorboards, predictions).
Users can choose to perform inference either on 1 data point or on his whole test dataset.
The inference scenario experiments main files only contain inference code.
[x] GeoLifeCLEF2023
The GLC23 dataset has been corrected and cleaned to better fit the sample data provided with it.
The code documentation has been improved.
[x] Dropped Python 3.8 support
Malpolon will no longer be tested on Python3.8 since 2024 will be the end of life of security updates for the language. For now Python 3.10 installation guides and testing are provided.
:memo: Changelog
[x] Scenario refactoring
“Ecologists” In this scenario, users possess a dataset of their own on which they want to train a model while being able to easily customize the training process and the model architecture.
“Inference” In this scenario, users possess both a test dataset and a trained model. They want to simply test the performances of their model with relevant metrics.
“Kaggle” In this scenario, users are participants of a GeoLifeCLEF challenge and look for plug-and-play examples of data loaders and/or training examples, specifically developed for the challenge dataset.
Updated config files with new paths and cleaned example files. All example files now use Hydra v1.3 which does not change the current working directory to
hydra.run.dir
, so all paths given to config files should be relative to the experiment directory (e.g.:malpolon/examples/ecologists/sentinel-2a/
).Malpolon's engine classes (in particular the datasets classes) have been split for better encapsulation.
Metrics calling have been updated. A new dictionary was added in
malpolon.data.utils
containing torchmetrics names as keys and functional metrics callable as values. Hence users only need to add an entry in their config file with the same name as the dict keys + the torchmetrics kwargs under it. If a user wants to use a torchmetrics functional metric not listed in the dictionary, they can choose a custom name as entry for the config file and need to a sub-entrycallable: <Fmetrics.X.X>
.[x] Inference Added inference computing methods and methods to output inference predictions (metrics, logs, tensorboards, predictions). Users can choose to perform inference either on 1 data point or on his whole test dataset.
The inference scenario experiments main files only contain inference code.
[x] GeoLifeCLEF2023 The GLC23 dataset has been corrected and cleaned to better fit the sample data provided with it. The code documentation has been improved.
[x] Dropped Python 3.8 support Malpolon will no longer be tested on Python3.8 since 2024 will be the end of life of security updates for the language. For now Python 3.10 installation guides and testing are provided.
:link: Links
40
39
37
:white_check_mark: Checklist