hankcs/HanLP · error · ValueError

Unrecognized type for {embed}

Error message

Unrecognized type for {embed}

What it means

CharRNN's constructor accepts embed either as an int (embedding dim, builds nn.Embedding internally) or as an nn.Module with an embedding_dim attribute (reuses it and reads its dim for LSTM input_size). Anything else (str, float, None) raises this error before the LSTM is constructed.

Source

Thrown at hanlp/layers/embeddings/char_rnn.py:40

        """Character level RNN embedding module.

        Args:
            field: The field in samples this encoder will work on.
            vocab_size: The size of character vocab.
            embed: An ``Embedding`` object or the feature size to create an ``Embedding`` object.
            hidden_size: The hidden size of RNNs.
        """
        super(CharRNN, self).__init__()
        self.field = field
        # the embedding layer
        if isinstance(embed, int):
            self.embed = nn.Embedding(num_embeddings=vocab_size,
                                      embedding_dim=embed)
        elif isinstance(embed, nn.Module):
            self.embed = embed
            embed = embed.embedding_dim
        else:
            raise ValueError(f'Unrecognized type for {embed}')
        # the lstm layer
        self.lstm = nn.LSTM(input_size=embed,
                            hidden_size=hidden_size,
                            batch_first=True,
                            bidirectional=True)

    def forward(self, batch, mask, **kwargs):
        x = batch[f'{self.field}_char_id']
        # [batch_size, seq_len, fix_len]
        mask = x.ne(0)
        # [batch_size, seq_len]
        lens = mask.sum(-1)
        char_mask = lens.gt(0)

        # [n, fix_len, n_embed]
        x = self.embed(batch) if isinstance(self.embed, EmbeddingDim) else self.embed(x[char_mask])
        x = pack_padded_sequence(x[char_mask], lens[char_mask].cpu(), True, False)
        x, (h, _) = self.lstm(x)

View on GitHub (pinned to ddb1299bdd)

Solutions

1. Pass an int dim or a module exposing .embedding_dim (e.g. nn.Embedding)
2. Coerce config values: int(embed) when it's a numeric string
3. Wrap custom embeddings in a small module that defines embedding_dim

Example fix

# before
embed = CharRNNEmbedding(vocab, embed='100')  # str
# after
embed = CharRNNEmbedding(vocab, embed=int('100'))

if isinstance(embed, str) and embed.isdigit():
    embed = int(embed)
assert isinstance(embed, int) or (isinstance(embed, nn.Module) and hasattr(embed, 'embedding_dim'))

def valid_char_rnn_embed(embed) -> bool:
    return isinstance(embed, int) or (isinstance(embed, nn.Module) and hasattr(embed, 'embedding_dim'))

When it happens

Trigger: Passing embed='100' (string from config), a float, or an nn.Module lacking embedding_dim (e.g. a raw Linear) to CharRNNEmbedding.

Common situations: Config files yielding strings; passing a pretrained embedding wrapper that doesn't expose embedding_dim; passing None due to a missing config key.

Related errors

• Unrecognized type for {embed}
• Unsupported parameter type: {embed}
• error
• output ({}) must be of type bool or str
• Call fit or load before evaluate.

AI-assisted analysis of hankcs/HanLP@ddb1299bdd (2026-08-27). Data as JSON: /api/errors/5a81d6ab76906a65. Report an issue: GitHub.