efficientnet_b0

orchard.architectures.efficientnet_b0

EfficientNet-B0 Architecture for 224x224 Image Classification.

Adapts EfficientNet-B0 (compound scaling architecture) for image classification with transfer learning support. Handles both RGB and grayscale inputs through dynamic first-layer adaptation.

Key Features:

• Efficient Scaling: Balances depth, width, and resolution
• Transfer Learning: Leverages ImageNet pretrained weights
• Adaptive Input: Customizes first layer for grayscale datasets
• Channel Compression: Weight morphing for RGB→grayscale (3→1) adaptation

build_efficientnet_b0(num_classes, in_channels, *, pretrained)

Constructs EfficientNet-B0 adapted for image classification datasets.

Workflow

1. Load pretrained weights from ImageNet (if enabled)
2. Modify first conv layer to accept custom input channels
3. Apply weight morphing for channel compression (if grayscale)
4. Replace classification head with dataset-specific linear layer

Parameters:

Name	Type	Description	Default
num_classes	int	Number of dataset classes for classification head	required
in_channels	int	Input channels (1=Grayscale, 3=RGB)	required
pretrained	bool	Whether to load ImageNet pretrained weights	required

Returns:

Type	Description
Module	Adapted EfficientNet-B0 model (device placement handled by factory).

Source code in orchard/architectures/efficientnet_b0.py

def build_efficientnet_b0(
    num_classes: int,
    in_channels: int,
    *,
    pretrained: bool,
) -> nn.Module:
    """
    Constructs EfficientNet-B0 adapted for image classification datasets.

    Workflow:
        1. Load pretrained weights from ImageNet (if enabled)
        2. Modify first conv layer to accept custom input channels
        3. Apply weight morphing for channel compression (if grayscale)
        4. Replace classification head with dataset-specific linear layer

    Args:
        num_classes: Number of dataset classes for classification head
        in_channels: Input channels (1=Grayscale, 3=RGB)
        pretrained: Whether to load ImageNet pretrained weights

    Returns:
        Adapted EfficientNet-B0 model (device placement handled by factory).
    """
    # --- Step 1: Initialize with Optional Pretraining ---
    weights = models.EfficientNet_B0_Weights.IMAGENET1K_V1 if pretrained else None
    model = models.efficientnet_b0(weights=weights)

    # Snapshot original first conv layer (before replacement)
    old_conv = model.features[0][0]

    # --- Step 2: Adapt First Convolutional Layer ---
    # EfficientNet expects 3-channel input; modify for grayscale if needed
    new_conv = nn.Conv2d(
        in_channels=in_channels,  # Custom: 1 or 3
        out_channels=32,  # EfficientNet standard
        kernel_size=(3, 3),
        stride=(2, 2),  # Original EfficientNet stem (matches pretrained spatial statistics)
        padding=(1, 1),
        bias=False,
    )

    # --- Step 3: Weight Morphing (Transfer Pretrained Knowledge) ---
    if pretrained:
        morph_conv_weights(old_conv, new_conv, in_channels)

    # Replace entry layer with adapted version
    model.features[0][0] = new_conv

    # --- Step 4: Modify Classification Head ---
    # Replace ImageNet 1000-class head with dataset-specific projection
    model.classifier[1] = nn.Linear(model.classifier[1].in_features, num_classes)  # 1280 features

    return cast(nn.Module, model)