util.FPN

class lucid.models.objdet.util.FPN(in_channels_list: list[int], out_channels: int = 256)

The FPN (Feature Pyramid Network) module constructs a multi-scale feature pyramid from a list of input feature maps. It enhances semantic representation at all scales, typically used in detection backbones.

Constructor

def __init__(in_channels_list: list[int], out_channels: int = 256) -> None

Parameters

  • in_channels_list (list[int]): List of channel dimensions for input feature maps (e.g., from backbone layers C2, C3, C4, C5).

  • out_channels (int, optional): Number of output channels per feature map in the pyramid. Defaults to 256.

Returns

  • FPN (nn.Module): A module that fuses multi-scale feature maps into semantically rich pyramidal outputs.

Forward Input & Output

def forward(features: list[Tensor]) -> list[Tensor]

  • features (list[Tensor]): Input feature maps from the backbone, typically of decreasing spatial size.

  • Returns (list[Tensor]): A list of feature maps of shape \((N, \text{out\_channels}, H_l, W_l)\) at each pyramid level, with lateral and top-down connections fused.

Note

  • Uses 1x1 lateral convs and 3x3 smoothing convs.

  • All outputs are aligned to the same number of channels (out_channels).

  • Maintains same resolution per level as the input list.

Example

>>> from lucid.models.objdet.util import FPN
>>> inputs = [lucid.random.randn(1, c, s, s) for c, s in zip([64, 128, 256, 512], [64, 32, 16, 8])]
>>> fpn = FPN(in_channels_list=[64, 128, 256, 512])
>>> outputs = fpn(inputs)
>>> for feat in outputs:
...     print(feat.shape)
(1, 256, 64, 64)
(1, 256, 32, 32)
(1, 256, 16, 16)
(1, 256, 8, 8)