DL model zoo for iota2
Now that iota2 has the ability to use Pytorch models for training and inference, it would be interesting to distribute pre-trained models in order to avoid re-training (for energy and time efficiency or lack of computational ressources). These models could be useful in 3 cases:
- To be used for inference. This assumes using the same kind of data as input and solving the same task (classification with the same nomenclature, regression of the same variable). This case would not perform training, but just inference.
- To be used as in the case above, but with fine-tuning during a training phase. This case would allow changing the prediction task (change in classification nomenclature, regression of a different variable).
- To be used as feature extractors, that is, transforming the input data before another ML model. In this case, the model is not retrained or fine-tuned. This case assumes, as the previous ones, the same type of input data.
Other use cases could be added.