opendatalab/MinerU · error · ValueError

PPLCNetV4 only supports 3 input channels, got {in_channels}.

Error message

PPLCNetV4 only supports 3 input channels, got {in_channels}.

What it means

Raised by PPLCNetV4.__init__ when in_channels is anything other than 3. The stem is hard-wired for RGB input, and both the det and rec v6 net configs assume 3-channel images, so grayscale/4-channel input is rejected up front.

Source

Thrown at mineru/model/utils/pytorchocr/modeling/backbones/rec_lcnetv4.py:291

    def forward(self, pixel_values):
        """返回四个 stage 的输出特征，供 det/rec 上层按需使用。"""
        hidden_states = self.convolution(pixel_values)
        feature_maps = []
        for block in self.blocks:
            hidden_states = block(hidden_states)
            feature_maps.append(hidden_states)
        return feature_maps


class PPLCNetV4(nn.Module):
    """PP-OCRv6 使用的 PPLCNetV4 backbone，支持 det small 和 rec small/medium。"""

    def __init__(self, det=False, model_size="small", in_channels=3, **kwargs):
        """按 det/rec 模式选择 v6 的固定网络配置。"""
        super().__init__()
        self.det = det
        if in_channels != 3:
            raise ValueError(f"PPLCNetV4 only supports 3 input channels, got {in_channels}.")
        config_dict = NET_CONFIG_DET if det else NET_CONFIG_REC
        if model_size not in config_dict:
            mode = "det" if det else "rec"
            raise ValueError(f"PPLCNetV4 {mode} model_size must be one of {list(config_dict)}, got {model_size}.")
        config = config_dict[model_size]
        self.encoder = PPLCNetV4Encoder(config["stem_channels"], config["block_configs"])
        stage_out_channels = [stage[-1][2] for stage in config["block_configs"]]
        self.out_channels = stage_out_channels if det else stage_out_channels[-1]

    def forward(self, x):
        """det 返回四级特征列表，rec 返回高度池化后的识别特征。"""
        feature_maps = self.encoder(x)
        if self.det:
            return feature_maps
        x = feature_maps[-1]
        if self.training:
            return F.adaptive_avg_pool2d(x, [1, 40])
        if x.shape[2] < 3:

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Convert inputs to 3-channel RGB before the model (cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)) and leave in_channels at the default 3.
2. Remove an explicit in_channels override from the config so the default is used.

Example fix

# before
PPLCNetV4(det=False, in_channels=1)

# after
img = cv2.cvtColor(gray_img, cv2.COLOR_GRAY2BGR)
PPLCNetV4(det=False)  # default in_channels=3

assert in_channels == 3, "PPLCNetV4 requires 3-channel RGB input; convert grayscale images before the model"

When it happens

Trigger: Building the PP-OCRv6 backbone with in_channels=1 (grayscale pipeline) or in_channels=4 (RGBA) via config or kwargs; a config key like in_channels: 1 inherited from another model.

Common situations: Adapting a document pipeline that pre-converts images to grayscale; merging configs from a different backbone that supported 1-channel input.

Related errors

• Unsupported activation: {name}
• PPLCNetV4 {mode} model_size must be one of {list(config_dict
• mode[{model_name}_model] is not implemented!
• mode[{model_name}_model] is not implemented!
• The function 'replace_sub()' is deprecated, please use 'upgr

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/42388d074b6c9b95. Report an issue: GitHub.