stelviopas / pp1cb_ss21 Goto Github PK

Right now, the splitting is done only on a protein level and without stratification.
This means that each fold contains roughly (except if the number of samples is not divisible by the number of folds) the same number of proteins. However, these folds are not stratified, meaning that it may be the case that one fold only consists of disordered proteins, and another one consists of structured proteins only.

Stratification might be achieved using sklearns StratifiedKFold, but we have to ensure that proteins are not split.

If noone has an idea how to do this, we could alternatively just implement a sanity check, for example plotting a z-score histogram for each fold.

In order to train a NN, we should load out the dataset into the PyTorch DataLoader. However, the length of X and y are not equal, so one has to pre-process or pad them.
To do so, we can define collate_fn function which rewrites the default alignment based on samples length.
This should be done in the disorer_dataset.py and passed as an argument to DataLoader() in the cross-validation loop.

More about collate_fn can be found here

https://discuss.pytorch.org/t/how-to-use-collate-fn/27181