PaddlePaddle/PaddleOCR · error · TypeError

The mixer must be one of [Global, Local, Conv]

Error message

The mixer must be one of [Global, Local, Conv]

What it means

SVTRNet's Block (mixer stage) supports three attention mixers: 'Global' and 'Local' (both routed to the Attention mixer, where Local uses windowed attention when HW is set) and 'Conv' (ConvMixer). Any other mixer string raises TypeError listing the three valid values. Note it is a TypeError raised from __init__ of each Block, so it fires while the SVTR backbone is being assembled.

Source

Thrown at ppocr/modeling/backbones/rec_svtrnet.py:254

            self.norm1 = eval(norm_layer)(dim, epsilon=epsilon)
        else:
            self.norm1 = norm_layer(dim)
        if mixer == "Global" or mixer == "Local":
            self.mixer = Attention(
                dim,
                num_heads=num_heads,
                mixer=mixer,
                HW=HW,
                local_k=local_mixer,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                attn_drop=attn_drop,
                proj_drop=drop,
            )
        elif mixer == "Conv":
            self.mixer = ConvMixer(dim, num_heads=num_heads, HW=HW, local_k=local_mixer)
        else:
            raise TypeError("The mixer must be one of [Global, Local, Conv]")

        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else Identity()
        if isinstance(norm_layer, str):
            self.norm2 = eval(norm_layer)(dim, epsilon=epsilon)
        else:
            self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp_ratio = mlp_ratio
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )
        self.prenorm = prenorm

    def forward(self, x):
        if self.prenorm:

View on GitHub (pinned to 2661c7c0ef)

Solutions

1. Use only 'Global', 'Local', or 'Conv' (exact casing) in the Mixer list
2. If following an SVTRv2-style recipe, use the config/backbone file that implements those mixers instead of rec_svtrnet.py's Block
3. Verify each Mixer entry against a shipped PP-OCR rec SVTR config before editing

Example fix

# before
Mixer: ['Conv', 'Conv', 'Globalx']   # typo -> TypeError

# after
Mixer: ['Conv', 'Conv', 'Global']

MIXERS = {'Global', 'Local', 'Conv'}
assert all(m in MIXERS for m in Mixer), f'Mixer entries must be in {sorted(MIXERS)}, got {Mixer}'

def valid_mixers(mixer_list) -> bool:
    s = {'Global', 'Local', 'Conv'}
    return isinstance(mixer_list, (list, tuple)) and all(m in s for m in mixer_list)

When it happens

Trigger: SVTR config with Mixer=['Global','Local','Convx'] (typo), Mixer=['global'] (case-sensitive), or a name from a newer SVTR variant not implemented in this file (e.g. 'Partial' from SVTRv2-style configs).

Common situations: Porting an SVTRv2 / PP-OCRv4+ rec config whose mixer names differ from the classic SVTR implementation; case or spelling drift in hand-written Mixer lists.

Related errors

• mode[{mode}_model] is not implemented!
• {} is not supported in MultiLoss yet
• mode[{model_name}_model] is not implemented!
• RecResizeImg.image_shape is required in rec inference.yml
• Unexpected recognition channels: ${String(channels)}

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/da97f4e3eedd253c. Report an issue: GitHub.