4. Model Architecture

BreedClassifier (src/model.py)

class BreedClassifier(nn.Module):
    backbone: EfficientNetLite     # stem → 7 MBConv stages → head(1280ch)
    attention: CBAM | SEBlock      # inserted after stage 3 (88ch lite2 / 112ch lite4)
    avg_pool: AdaptiveAvgPool2d(1)
    binary_head: Linear(1280→256→2)
    cattle_head: Linear(1280→512→57) + Dropout(0.3)
    buffalo_head: Linear(1280→512→18) + Dropout(0.3)

Forward Path

1. forward_features(x): backbone stages 0..3 → CBAM → stages 4..6 → head → pool → flatten
2. forward(x): features → 3 parallel heads → dict{binary, cattle, buffalo, features}

Freeze/Unfreeze Methods

• freeze_backbone(): freezes backbone + attention
• freeze_all() / unfreeze_all(): all parameters
• backbone_eval() / backbone_train(): BatchNorm mode control

EfficientNet-Lite (src/efficientnet_lite.py)

• MBConv blocks with relu6 activation (no SE in lite variant)
• forward_until(x, stage_idx) / forward_from(x, stage_idx): split forward for attention insertion
• Weight loading: load_backbone_weights() — strict=False, allows missing fc.*

Attention (src/cbam.py)

• CBAM: ChannelAttention(avg+max pool → MLP → sigmoid) → SpatialAttention(avg+max concat → conv → sigmoid)
• SE: AdaptiveAvgPool → FC reduce → ReLU → FC expand → Sigmoid

---