Christiaan Meijer


we want to completely get rid of the tutorial repo.

Don't keep all models in memory during training

5

Make sure that only the current model is loaded, and the rest is stored somewhere else.

tutorial to tensorflow

1

check to see if tutorial is already using TF2 instead of Keras. Make sure it does.

Create abstract base class for models

1

This abc can handle the common hyperparameters like lr and reg. It should also have a wrapper function like this: ``` def generate_model: p = self.generate_hyperparameters() return create_model(p) ```

model_comparison.json contains multiple metrics. The visualization does not support selecting any of these metrics to show yet, and now shows the accuracy by default (it breaks when accuracy is not...

Modelgen.generate_models() is agnostic about all the hyperparameters of the possible models. It's docstring therefore does not list them. There should be a link there, to some documentation where the user...

https://github.com/NLeSC/mcfly/blob/5ebeefc392c777372f943dfae51f7fcc48e4123d/tests/test_find_architecture.py#L61 This test should be removed or fixed.

Our tests now only test the case that 'accuracy' is the metric to base the choice of the best model on.