morphoclass.models.bidirectional_net module

Implementation of the BidirectionalNet classifier.

class morphoclass.models.bidirectional_net.BidirectionalNet(num_classes, num_nodes_features)

Bases: torch.nn.modules.module.Module

Model for classifying morphologies of pyramidal neurons.

This is the architecture that performed best in the TensorFlow implementation. It consists of two graph convolutions layers with each computing two convolutions: one on the directed adjacency matrix, and one with the adjacency matrix with the reversed direction. The results of both convolutions are concatenated and passed to the next layer. After the two parallel graph convolutions follows a global average pooling layer and a fully-connected layer. Finally, a softmax layer is used for prediction.

Parameters

  • num_classes (int) – The number of output classes.

  • num_nodes_features (int) – The number of input node features.

accuracy(data)

Run the forward pass and compute the accuracy.

Parameters

data (torch_geometric.data.Data) – The input batch of data.

Returns

acc – The accuracy on the current data batch.

Return type

float

forward(data)

Compute the forward pass.

Parameters

data (torch_geometric.data.data.Data) – A batch of input data.

Returns

The log softmax of the predictions.

Return type

log_softmax

loss_acc(data)

Run the forward pass and compute the loss and accuracy.

Parameters

data (torch_geometric.data.Data) – The input batch of data.

Returns

  • loss (float) – The loss on the given data batch.

  • acc (float) – The accuracy on the current data batch.

training: bool