Skip to content

lignn

LogicLoss

Bases: Module

Theory-informed loss function for determining topology diagram validity

Source code in src/torch_structure/loss/lignn.py 
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
class LogicLoss(torch.nn.Module):
    """Theory-informed loss function for determining topology diagram validity"""

    def __init__(self, search_depth, device, reduction="mean"):
        super(LogicLoss, self).__init__()
        self.search_depth = search_depth
        self.device = device
        self.local = LocalLogicLoss(reduction=reduction)
        self.wmax = WeightedMax()
        self.nllloss = NLLLoss(reduction=reduction)
        self.mean_dim = (0, 1) if reduction == "mean" else 1

    def forward(self, x, edge_index, batch, y, x2):
        """Compute the combined local and trail-connectivity loss.

        Args:
            x (torch.Tensor [N, C]): per-node predicted probability channels.
            edge_index (torch.Tensor [2, E]): edge connectivity.
            batch (torch.Tensor [N]): unused, kept for interface consistency.
            y (torch.Tensor [N]): target trail-edge count per node, used by
                the local loss.
            x2 (torch.Tensor [N, C]): initial values propagated over
                ``search_depth`` hops to assess indirect trail connectivity.

        Returns:
            tuple[torch.Tensor, torch.Tensor, torch.Tensor]: the combined
            loss, the local loss, and the trail loss.
        """
        # Compress domain to prevent numerical issues with ones and zeros
        x = x * 0.9999999 + 0.00000005

        # Compute local loss related to local trail connectivity
        local_loss = self.local(x, edge_index, y, self.device)

        # Compute trail loss related to indirect connectivity of nodes to support/origin nodes through trail edges
        for i in range(self.search_depth):
            x2 = self.wmax(x2, edge_index, weight=x[:, 1])
        trail_loss = torch.mean(-torch.log(x2), dim=self.mean_dim)

        # Compute weighted average according to local_component
        local_component = 0.9
        loss = local_component * local_loss + (1 - local_component) * trail_loss
        return loss, local_loss, trail_loss

forward(x, edge_index, batch, y, x2)

Compute the combined local and trail-connectivity loss.

Parameters:

Name Type Description Default
x Tensor[N, C]

per-node predicted probability channels.

 required
edge_index Tensor[2, E]

edge connectivity.

 required
batch Tensor[N]

unused, kept for interface consistency.

 required
y Tensor[N]

target trail-edge count per node, used by the local loss.

 required
x2 Tensor[N, C]

initial values propagated over search_depth hops to assess indirect trail connectivity.

 required

Returns:

Type Description

tuple[torch.Tensor, torch.Tensor, torch.Tensor]: the combined

loss, the local loss, and the trail loss.
Source code in src/torch_structure/loss/lignn.py 
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
def forward(self, x, edge_index, batch, y, x2):
    """Compute the combined local and trail-connectivity loss.

    Args:
        x (torch.Tensor [N, C]): per-node predicted probability channels.
        edge_index (torch.Tensor [2, E]): edge connectivity.
        batch (torch.Tensor [N]): unused, kept for interface consistency.
        y (torch.Tensor [N]): target trail-edge count per node, used by
            the local loss.
        x2 (torch.Tensor [N, C]): initial values propagated over
            ``search_depth`` hops to assess indirect trail connectivity.

    Returns:
        tuple[torch.Tensor, torch.Tensor, torch.Tensor]: the combined
        loss, the local loss, and the trail loss.
    """
    # Compress domain to prevent numerical issues with ones and zeros
    x = x * 0.9999999 + 0.00000005

    # Compute local loss related to local trail connectivity
    local_loss = self.local(x, edge_index, y, self.device)

    # Compute trail loss related to indirect connectivity of nodes to support/origin nodes through trail edges
    for i in range(self.search_depth):
        x2 = self.wmax(x2, edge_index, weight=x[:, 1])
    trail_loss = torch.mean(-torch.log(x2), dim=self.mean_dim)

    # Compute weighted average according to local_component
    local_component = 0.9
    loss = local_component * local_loss + (1 - local_component) * trail_loss
    return loss, local_loss, trail_loss