This sections covers how to get started with ConfigILM.

There are currently two ways of installing ConfigILM. The recommended way is using pypi via

!pip install configilm

Alternatively, the source can be directly downloaded from github. If you want to use the extras (e.g. Datasets), installation is dependent on which extras you want to use. pip install configilm[bigearthnet] installs all additional dependencies for the BigEarthNet dataset loading using LMDB and the BigEarthNet-Encoder. For the DataModules and use with PyTorch Lightning, use pip install configilm[full-lightning] (same as pip install configilm[full]) for use with lightning.pytorch and pip install configilm[full-pytorch_lightning] for use with pytorch_lightning. Both versions also include the dependencies of BigEarthNet. As both versions are equal in their features, the choice is up to the user or the use case. For example, TorchGeo uses lightning.pytorch, so it would be better to use pip install configilm[full-lightning] in combination with TorchGeo. For simplicity, the full version is used in the examples in this documentation.


Using the extras means installing some additional dependencies.


If you use the wheel, installation would equate to pip install configilm-0.4.10-py3-none-any.whl and pip install configilm-0.4.10-py3-none-any.whl[extra]

ConfigILM allows you to easily combine and use predefined image and language models and use them for tasks such as Supervised Pretraining or VQA. For this, models can be trained in an end-to-end fashion or pre-trained checkpoints can be used. How the configuration is set up is explained on the next page. Alternatively, use cases for this framework are presented on the following pages.

For all examples in this documentation it is assumed that ConfigILM has been installed together with the [extra] dependencies.

If you want to try out the framework’s functionality, check out the google colab, which shows an example use case.