opendatalab/MinerU · error · ValueError

Unsupported activation: {name}

Error message

Unsupported activation: {name}

What it means

Raised by _build_activation in the PP-LCNetV4 recognition backbone when a block config names an activation that the factory cannot map. Supported names are relu, gelu, silu/swish, hardsigmoid (plus None -> Identity). It exists so HF/Paddle config strings map cleanly onto parameterless torch nn layers.

Source

Thrown at mineru/model/utils/pytorchocr/modeling/backbones/rec_lcnetv4.py:76

            [[3, 512, 768, (2, 1), False], [3, 768, 768, 1, True], [3, 768, 768, 1, False]],
        ],
    },
}


def _build_activation(name):
    """按 PP-OCRv6 配置创建无参数激活层，便于和 safetensors 权重命名保持解耦。"""
    if name is None:
        return nn.Identity()
    if name == "relu":
        return nn.ReLU()
    if name == "gelu":
        return nn.GELU()
    if name in {"silu", "swish"}:
        return nn.SiLU()
    if name == "hardsigmoid":
        return nn.Hardsigmoid()
    raise ValueError(f"Unsupported activation: {name}")


def _to_stride(stride):
    """把 Paddle/Transformers 配置里的 stride 统一成 PyTorch 可接受的格式。"""
    if isinstance(stride, list):
        return tuple(stride)
    return stride


class PPLCNetV4ConvLayer(nn.Module):
    """PP-LCNetV4 的 Conv-BN-Act 基础层，属性名对齐 HF 权重。"""

    def __init__(
        self,
        in_channels,
        out_channels,
        kernel_size,
        stride=1,

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Use only the supported activation names: relu, gelu, silu (or swish), hardsigmoid, or None.
2. If you need hard_swish, map it explicitly: hard_swish == hardsigmoid-family activation — extend _build_activation with nn.Hardswish if you control the fork.
3. Restore the shipped NET_CONFIG_REC/NET_CONFIG_DET tables if they were edited.

Example fix

# before (fork edit)
if name == "hard_swish": ...
# config uses "hard_swish" -> raises

# after
_build_activation accepts "hard_swish":
    if name in {"hardsigmoid", "hard_sigmoid"}:
        return nn.Hardsigmoid()
    if name == "hard_swish":
        return nn.Hardswish()

SUPPORTED_ACT = {None, "relu", "gelu", "silu", "swish", "hardsigmoid"}
assert act_name in SUPPORTED_ACT, f"unsupported activation: {act_name!r}"

def is_supported_activation(name) -> bool:
    return name in {None, "relu", "gelu", "silu", "swish", "hardsigmoid"}

When it happens

Trigger: Instantiating PPLCNetV4 backbone with a hand-modified NET_CONFIG (or passing config dicts) that use strings like 'hard_swish', 'relu6', 'sigmoid', or 'GELU' (case-sensitive).

Common situations: Porting LCNet configs from PaddleOCR/HF where 'hard_swish' is standard; editing the v4 block tables to experiment; case mismatches.

Related errors

• PPLCNetV4 {mode} model_size must be one of {list(config_dict
• PPLCNetV4 only supports 3 input channels, got {in_channels}.
• mode[{model_name}_model] is not implemented!
• mode[{model_name}_model] is not implemented!
• The mixer must be one of [Global, Local, Conv]

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/09fc26a1387555a0. Report an issue: GitHub.