There are some issues when using the package for testing a pre-trained model on newly generated data:

• The GraphDataset class requires dataset_train as input even in such cases (whenever train is False). We should be able of using a test dataset without the need of the original model's training dataset. We can use the info stored in the pre-trained model for inherit the needed attributes. (see _check_inherited_params in dataset.py)
• In the Trainer class' init, before loading parameters and the pretrained model there is a check for the target, which in a pre-trained model case could be not present at all.
• The Trainer class expects the attribute epoch_saved_model, which should be saved within the state of the pre-trained model.
• If the test dataset has no labels, the output exporter doesn't work (ValueError("All arrays must be of the same length"))

In order to make reasonable changes, I think we need to take into account all the possible scenarios using a mock example:

• No pre-trained model, train, valid, and test. (should be good)
• No pre-trained model, train, valid, no test. (should be good)
• No pre-trained model, train only. (should be good)
• Pre-trained model, test only, with labels. (the one to improve the code for)
• Pre-trained model, test only, with no labels. (the one to improve the code for)