RCNNConfig¶

class lucid.models.RCNNConfig(backbone: lucid.nn.module.Module, feat_dim: int, num_classes: int, image_means: tuple[float, float, float] | list[float] = (0.485, 0.456, 0.406), pixel_scale: float = 1.0, warper_output_size: tuple[int, int] | list[int] = (224, 224), nms_iou_thresh: float = 0.3, score_thresh: float = 0.0, add_one: bool = True)¶

RCNNConfig stores the backbone, classifier width, input normalization, region warping size, and post-processing thresholds used by lucid.models.RCNN.

Class Signature¶

@dataclass
class RCNNConfig:
    backbone: nn.Module
    feat_dim: int
    num_classes: int
    image_means: tuple[float, float, float] | list[float] = (0.485, 0.456, 0.406)
    pixel_scale: float = 1.0
    warper_output_size: tuple[int, int] | list[int] = (224, 224)
    nms_iou_thresh: float = 0.3
    score_thresh: float = 0.0
    add_one: bool = True

Parameters¶

backbone (nn.Module): Shared region feature extractor.
feat_dim (int): Flattened feature width emitted by the backbone.
num_classes (int): Number of classification logits predicted per region.
image_means: Per-channel image normalization means.
pixel_scale (float): Input pixel scaling factor applied before mean subtraction.
warper_output_size: Fixed crop resolution used by the region warper.
nms_iou_thresh (float): IoU threshold used during non-maximum suppression.
score_thresh (float): Minimum class probability kept during prediction.
add_one (bool): Whether bounding-box delta decoding uses the +1 width/height convention.

Validation¶

backbone must be an nn.Module.
feat_dim and num_classes must be greater than 0.
image_means must contain exactly three values.
pixel_scale must be greater than 0.
warper_output_size must contain exactly two positive integers.
nms_iou_thresh must be in [0, 1].

Usage¶

import lucid.models as models
import lucid.nn as nn

backbone = nn.Sequential(
    nn.Conv2d(3, 64, kernel_size=3, padding=1),
    nn.ReLU(),
    nn.AdaptiveAvgPool2d((1, 1)),
)
config = models.RCNNConfig(backbone=backbone, feat_dim=64, num_classes=3)
model = models.RCNN(config)