What causes "illegal stride value." in WZMIAOMIAO/deep-learning-for-image-processing?

WZMIAOMIAO/deep-learning-for-image-processing · error · ValueError

illegal stride value.

Error message

illegal stride value.

What it means

MBConv only supports strides 1 or 2, since shortcut downsampling and depthwise conv are built around those two values. Any other stride is rejected at module construction.

Source

Thrown at pytorch_classification/Test11_efficientnetV2/model.py:112

        scale = self.conv_expand(scale)
        scale = self.act2(scale)
        return scale * x


class MBConv(nn.Module):
    def __init__(self,
                 kernel_size: int,
                 input_c: int,
                 out_c: int,
                 expand_ratio: int,
                 stride: int,
                 se_ratio: float,
                 drop_rate: float,
                 norm_layer: Callable[..., nn.Module]):
        super(MBConv, self).__init__()

        if stride not in [1, 2]:
            raise ValueError("illegal stride value.")

        self.has_shortcut = (stride == 1 and input_c == out_c)

        activation_layer = nn.SiLU  # alias Swish
        expanded_c = input_c * expand_ratio

        # 在EfficientNetV2中，MBConv中不存在expansion=1的情况所以conv_pw肯定存在
        assert expand_ratio != 1
        # Point-wise expansion
        self.expand_conv = ConvBNAct(input_c,
                                     expanded_c,
                                     kernel_size=1,
                                     norm_layer=norm_layer,
                                     activation_layer=activation_layer)

        # Depth-wise convolution
        self.dwconv = ConvBNAct(expanded_c,
                                expanded_c,

View on GitHub (pinned to 1ec3fe6f37)

Solutions

Use stride=1 or stride=2 only.
To downsample more, stack multiple stride-2 blocks instead of one large-stride block.
Fix the cfg/stage definition supplying the bad stride value.

Example fix

// before
MBConv(input_c=24, out_c=48, stride=3, ...)
// after
MBConv(input_c=24, out_c=48, stride=2, ...)  # chain two stride-2 blocks for /4 downsampling

Defensive patterns

Strategy: validation

Validate before calling

def build_mbconv(cfg):
    stride = cfg["stride"]
    assert stride in (1, 2), f"MBConv stride must be 1 or 2, got {stride}"
    return MBConv(input_c=cfg["in"], out_c=cfg["out"], stride=stride, ...)

Type guard

def is_legal_stride(s) -> bool:
    return s in (1, 2)

Try / catch

try:
    block = MBConv(input_c, out_c, stride=s, ...)
except ValueError as e:
    print(e)
    block = MBConv(input_c, out_c, stride=min(s, 2), ...)

Prevention

Only use strides 1 or 2 in stage configs
Achieve larger downsampling by stacking stride-2 blocks
Validate cfg stride fields before constructing the model

When it happens

Trigger: Constructing MBConv with stride=3 or 0, typically from a hand-edited cfg where 'stride' was changed, or a custom stage list passing a wrong stride.

Common situations: Editing the EfficientNetV2 cfg dict to increase downsampling (someone sets stride=3), or copying a block spec from another network with larger strides.

Understand the failure class

Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.

Related errors

AI-assisted analysis of WZMIAOMIAO/deep-learning-for-image-processing@1ec3fe6f37 (2026-08-30). Data as JSON: /api/errors/1c9e07ec52116be3. Report an issue: GitHub.