Properly handle pre-trained ML models for LR

[AKuederle]

Currently we provide pre-trained Sklearn models (#52) for LR detection. They are stored as pickle files. This is an issue, as future versions of sklearn might not be able to load the files.

Ideally we should create a script that can retrain these models from scratch. However, the data the models are originally based on is not public.

• Identify which models we actually need for the final pipelines -> Only MS_project_all_all
• Identify the Data -> MSProject data available (internally) on Escience -> Algorothm Test Study -> data -> USFD
• Create dataset class for MSProject dataset
• Create CV and retraining script

Original model was trained with following parameters (optimized in a GridsearchCV (n_folds=5, optimized for accuracy), Model was a SVC with MinMaxScaler

param_grid_svm_lin = [
    {'C': [0.01, 0.1, 1, 10, 100, 1000, 10000], 'kernel': ['linear']}]

For the dataset:

• There is a clinical info file on EScience which could be the basis for the index.
• Further information from the original code

if "msproject" in dataset["name"]:
        hc_list1 = ["C"+str(i) for i in range(1,9)] 
        hc_list2 = ["MS019","MS027","MS036","MS049","MS050","MS051","MS052","MS053","MS056","MS057","MS058","MS059","MS060","MS063","MS064","MS065"]
        if "all" in dataset["name"]:
            patient_list = patient_list
        elif "hc" in dataset["name"]:
            # not only folders with C belong to controls (referring to clinical info from e-science)
            patient_list = np.concatenate([hc_list1, hc_list2])
        else:
            hc_list = np.concatenate([hc_list1, hc_list2])
            patient_list = np.setdiff1d(patient_list,hc_list)

# exclude certain subjects due to data errors

if "MS021" in train_ids.values:
    train_ids = train_ids.loc[train_ids != "MS021"]
if "MS025" in train_ids.values:
    train_ids = train_ids.loc[train_ids != "MS025"]

Further notes:

Eventhought that problem could be completely solved using sklearn, it might be helpful to use the respective tpcp tooling to be able to use the pipelines we have and also create the basis for L/R validation for any algorithm

[AKuederle]

Notes:

• We should keep the oldmodels around (unless newly trained are really identical -> unlikely)
• The full retraining code should be kept around as a no_exc example -> Good documentation.