Skip-Connection Level Spacing

level_gap controls the vertical distance in pixels between overlapping skip-connection routes. This nested residual model creates two detour levels: an inner shortcut around one convolution and an outer shortcut around the whole block. Compare a compact setting with a more widely spaced one. The additional rendering options separate the boxes and keep attention on the routed lines.

import matplotlib.pyplot as plt
import torch
import visualtorch
from torch import nn


class NestedResidualBlock(nn.Module):
    """A residual block with one shortcut nested inside another."""

    def __init__(self, channels: int) -> None:
        super().__init__()
        self.conv1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
        self.conv2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
        self.inner_relu = nn.ReLU()
        self.conv3 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
        self.outer_relu = nn.ReLU()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Run the block with overlapping inner and outer shortcuts."""
        outer_identity = x
        x = self.conv1(x)
        inner_identity = x
        x = self.inner_relu(self.conv2(x) + inner_identity)
        x = self.conv3(x)
        return self.outer_relu(x + outer_identity)


model = NestedResidualBlock(channels=4)
input_shape = (1, 4, 16, 16)

level_gaps = (20, 80)
images = [
    visualtorch.render(
        model,
        input_shape,
        style="lenet",
        level_gap=level_gap,
        spacing=25,
        show_dimension=False,
        connector_width=2,
    )
    for level_gap in level_gaps
]

dpi = 150  # rendered at 2x this in the final doc build (savefig.dpi=300 in conf.py)
figure_width = min(12, max(8, sum(img.width for img in images) / dpi))
figure_height = min(4, max(3, max(img.height for img in images) / dpi))
_, axes = plt.subplots(1, 2, figsize=(figure_width, figure_height), dpi=dpi)
for ax, img, level_gap in zip(axes, images, level_gaps, strict=True):
    ax.imshow(img)
    ax.set_title(f"level_gap={level_gap}")
    ax.axis("off")

plt.tight_layout()
plt.show()