Please only change inline documentation scince #21

Paper: Run PROTEINS analysis

• Set up PROTEINS analysis
• Run PROTEINS analysis
• Make figure for PROTEINS analysis
• Select fMRI images with sex and age (decide)
• Run GCN analysis for sex
• Run GCN analysis for age
• Make figure for sex and age

The setup.py still needs old dependencies.

Update according to requirements.txt

After #53 is merged.

If the original GEM package will be made avaiable via pypi, we should switch to using it.

Adding GEM has been suggested in palash1992/GEM#106.

All possible export formats have to be tested.
YAML was removed recently because it did not work as expected.

The dot export format is adding a node when reading the generated files.
Please check every output method and remove the export formats that are not working.

load_conn function gets a simple test

This looks like a major construction site.

• Check that the conversion works as expected
• Why would anyone convert the graphs into this format without additional estimators / transformers handling this format
• Can this format simply be embedded into numpy arrays?

quick note:
dgl should be added to requirements.txt

Run the GraphPLithmusTest to evaluate the performance of very GCN arrchitecture on a simple classification task.

• Add test for sporadically failing graphs
  In a list of graphs and by extracting a list of metrics construct a graph in that list where a specific metric will fail to compute and make sure, that the output is as desired.

• Add tests for extract_measures

GraphConstructorSpatial requires a atlas_name_coords.csv. Please provide an example.
Is this constructor intedet for DTI or RestingState data?
Why is the default atlas HO not tested? Why is it not contained in the graph package?

Repair current allow_zero_in_degree option for GCNs (currently fails)

Should be done with #17

Is your feature request related to a problem? Please describe.
PHOTONAI-Graph only supports dgl graphs with the same number of node features. If you use dgl graphs with varying numbers of nodes/features (eg, dgl mini dataset), the pipeline does not work.

Describe the solution you'd like
Implement a new architecture that uses an MLP to sample down the number of node features to a fixed size. Also build a check that allows for using node degree from within the different architectures.

Additional context
This will be important for version 2.0

Unit Tests
Reimplement these unit tests that use the dgl mini dataset

    def test_gcn_classifier_dgl(self):
        gat_clf = GCNClassifierModel(nn_epochs=20)
        gat_clf.fit(self.X_dgl, self.y)
        output = gat_clf.predict(self.X_dgl)
        self.assertTrue(np.array_equal(np.array(output.shape), self.y.shape))

    def test_gat_classifier_dgl(self):
        gat_clf = GATClassifierModel(nn_epochs=20)
        gat_clf.fit(self.X_dgl, self.y)
        output = gat_clf.predict(self.X_dgl)
        self.assertEqual(output.shape, self.y.shape)

Lines 83 and following of GraphConstructor are:

    def prep_mtrx(self, adjacency):
        """transforms the matrix according to selected criteria"""
        if self.fisher_transform:
            adjacency = individual_fishertransform(adjacency)
        if self.use_abs:
            adjacency = np.abs(adjacency)
        if self.zscore:
            adjacency = individual_ztransform(adjacency)
        if self.use_abs_zscore:
            adjacency = np.abs(adjacency)

        return adjacency

• Does the order of this checks matter?
• Why is the fisher_transform not calculated on the absolute values of the adjacency if np.abs is passed
• Should there be a use_abs_fisher_transform paramter?
• Write Tests for all possible combinations of parameters

Why does getting started require such an advanced method for data loading?

Couldn't we simplify this to use for example nx.load (or some other photonai_graph based loader)?

This advanced matlab data loading could be used as a more advanced example.

NeuralNetworks should be revised.

dgl has a new GATConv layer in its newest version. Integrate the GATConv into the GATClassifier/Regressor and get rid of the GATLayer.