sage.architecture.backend.resnet3d_cbam

Filename : resnet3d_cbam.py Description : Short description of the file

Created on 2026-03-05 13:52:22

__author__ = Narenraju Nagarajan __copyright__ = Copyright 2026, Sage __license__ = MIT Licence __version__ = 0.0.1 __maintainer__ = Narenraju Nagarajan __affiliation__ = N/A __email__ = N/A __status__ = [‘inProgress’, ‘Archived’, ‘inUsage’, ‘Debugging’]

GitHub Repository: NULL

Documentation: NULL

Classes

ResNet

3D ResNet backbone with CBAM + detector-aware attention modules.

Functions

`resnet18_cbam`([pretrained])	Constructs a ResNet-18 model.
`resnet34_cbam`([pretrained])	Constructs a ResNet-34 model.
`resnet50_cbam`([pretrained])	Constructs a ResNet-50 model.
`resnet101_cbam`([pretrained])	Constructs a ResNet-101 model.
`resnet152_cbam`([pretrained])	Constructs a ResNet-152 model.

Module Contents

class ResNet(block, layers, num_classes=512, in_channels=1)[source]

Bases: torch.nn.Module

3D ResNet backbone with CBAM + detector-aware attention modules.

An extension of the standard 2D ResNet that processes detector feature volumes of shape (B, 1, D=2, H, W) with 3D convolutions. In addition to the per-block CBAM gates, the backbone integrates:

DetectorAttention in every residual block to weight each detector’s contribution channel-wise.
DetectorFusion after layer 2 to mix H1/L1 features.
CoherenceAttention after layer 2 to gate on inter-detector agreement.
TemporalTransformer after layer 4 for long-range context over the frequency axis.
Learned detector embedding added to the stem output for positional discrimination of the two detector channels.

Parameters:

block (type) – Residual block class — BasicBlock or Bottleneck.
layers (list[int]) – Number of blocks in each of the four stages.
num_classes (int) – Output feature dimension of the final FC layer (default 512).
in_channels (int) – Number of input channels (default 1 for single-channel 3D volume).
state (Initialize internal Module)
ScriptModule. (shared by both nn.Module and)