resnet_18

orchard.architectures.resnet_18

ResNet-18 Multi-Resolution Architecture.

Adaptive ResNet-18 supporting 28x28, 32x32, 64x64, 128x128, and 224x224 resolutions with resolution-specific architectural modifications.

Resolution-Specific Adaptations:

28x28 / 32x32:

• 7x7 Conv1 → 3x3 Conv1 (stride 1 instead of 2)
• MaxPool removed (prevents excessive spatial loss)
• Weight morphing via bicubic interpolation

64x64 / 128x128 / 224x224:

• Standard ResNet-18 stem (7x7 Conv1, stride 2, MaxPool)
• Only channel adaptation for grayscale inputs

build_resnet_18(num_classes, in_channels, *, pretrained, resolution)

Constructs ResNet-18 with resolution-aware architectural adaptation.

At 28x28/32x32, performs stem surgery to preserve spatial resolution. At 64x64, 128x128, and 224x224, uses the standard ResNet-18 architecture.

Workflow

1. Load ImageNet pretrained ResNet-18 (if enabled)
2. Apply resolution-specific stem adaptation
3. Replace classification head with dataset-specific linear layer

Parameters:

Name	Type	Description	Default
num_classes	int	Number of dataset classes	required
in_channels	int	Input channels (1=Grayscale, 3=RGB)	required
pretrained	bool	Whether to load ImageNet pretrained weights	required
resolution	int	Input image resolution (28, 32, 64, 128, or 224)	required

Returns:

Type	Description
Module	Adapted ResNet-18 (device placement handled by factory).

Source code in orchard/architectures/resnet_18.py

def build_resnet_18(
    num_classes: int,
    in_channels: int,
    *,
    pretrained: bool,
    resolution: int,
) -> nn.Module:
    """
    Constructs ResNet-18 with resolution-aware architectural adaptation.

    At 28x28/32x32, performs stem surgery to preserve spatial resolution.
    At 64x64, 128x128, and 224x224, uses the standard ResNet-18 architecture.

    Workflow:
        1. Load ImageNet pretrained ResNet-18 (if enabled)
        2. Apply resolution-specific stem adaptation
        3. Replace classification head with dataset-specific linear layer

    Args:
        num_classes: Number of dataset classes
        in_channels: Input channels (1=Grayscale, 3=RGB)
        pretrained: Whether to load ImageNet pretrained weights
        resolution: Input image resolution (28, 32, 64, 128, or 224)

    Returns:
        Adapted ResNet-18 (device placement handled by factory).
    """
    # --- Step 1: Initialize with Optional Pretraining ---
    weights = models.ResNet18_Weights.IMAGENET1K_V1 if pretrained else None
    model = models.resnet18(weights=weights)

    if resolution <= 32:  # 28, 32: small stem (3x3 stride-1, no MaxPool)
        _adapt_stem_28(model, in_channels, pretrained)
    else:  # 64, 224 — standard stem
        _adapt_stem_standard(model, in_channels, pretrained)

    # --- Step 3: Replace Classification Head ---
    model.fc = nn.Linear(model.fc.in_features, num_classes)

    return cast(nn.Module, model)