Global Attention Mechanism (GAM) Overview

The Global Attention Mechanism enhances feature discrimination by jointly modeling channel-wise and spatial dependencies. It employs a dual-path design: a multilayer perceptron (MLP) branch for channel attention and a lightweight convolutional branch for spatial attention, both operating on the same input tensor.

GAM Implemantation

import torch
import torch.nn as nn

class GlobalAttentionModule(nn.Module):
    def __init__(self, channels: int, reduction_ratio: int = 4):
        super().__init__()
        mid_channels = channels // reduction_ratio
        
        # Channel attention pathway: MLP with ReLU activation
        self.channel_proj = nn.Sequential(
            nn.Linear(channels, mid_channels),
            nn.ReLU(inplace=True),
            nn.Linear(mid_channels, channels)
        )
        
        # Spatial attention pathway: two 7×7 convolutions with batch norm
        self.spatial_proj = nn.Sequential(
            nn.Conv2d(channels, mid_channels, kernel_size=7, padding=3, bias=False),
            nn.BatchNorm2d(mid_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(mid_channels, channels, kernel_size=7, padding=3, bias=False),
            nn.BatchNorm2d(channels)
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        batch_size, channels, height, width = x.shape
        
        # Channel attention: reshape → MLP → sigmoid
        permuted = x.permute(0, 2, 3, 1).reshape(batch_size, -1, channels)
        channel_weights = self.channel_proj(permuted).reshape(batch_size, height, width, channels)
        channel_att = channel_weights.permute(0, 3, 1, 2).sigmoid()
        x = x * channel_att
        
        # Spatial attention: conv → sigmoid
        spatial_att = self.spatial_proj(x).sigmoid()
        return x * spatial_att

Integration Steps

1. Module Registration

Place GlobalAttentionModule in models/newmodels/gam.py, then expose it in models/newmodels/__init__.py:

from .gam import GlobalAttentionModule

2. Model Parser Extension

In models/yolo.py, add support in the parse_model function:

elif m is GlobalAttentionModule:
    args = [ch[f], *args]

Ensure GlobalAttentionModule is imported at the top of yolo.py.

3. YAML Configuration Update

Create models/detect/yolov9-gam.yaml. Insert the module at the final stage of the deteciton head — right before the detector layer. Adjust layer indices accordingly. Example excerpt:

# ... previous backbone and neck layers ...

# Final detection head layer
[-1, 1, RepNCSPELAN4, [512, 512, 256, 1]],  # 37
[-1, 1, GlobalAttentionModule, []],          # 38 ← inserted GAM block

# Detection layer: update indices to include new layer 38
[[31, 34, 38, 16, 19, 22], 1, DualDDetect, [nc]]

Note: All downstream references (e.g., in DualDDetect) must reflect updated indices if the new module shifts layer numbering.

4. Training Script Adjustment

Modify train_dual.py:

• Set --weights '' to train from scratch.
• Specify --cfg models/detect/yolov9-gam.yaml.
• Configure --data, --epochs, and --batch-size per hardware constraints.

Verification during training initialization confirms successful integration when the printed model architecture includes GlobalAttentionModule.