hankcs/HanLP · error · ValueError

activation must be callable: type={}

Error message

activation must be callable: type={}

What it means

The span-ranking SRL layer accepts an `activation` argument that must be callable (a function/nn.Module like torch.sigmoid or F.relu). Passing a string such as 'relu' or 'tanh' (a common convention in other libraries like sklearn/transformers) raises ValueError. None is allowed and means identity.

Source

Thrown at hanlp/components/srl/span_rank/layer.py:119

    def forward(self, x):
        if self.training:
            return torch.mul(x, self.drop_mask.to(x.device))
        else:  # eval
            return x * (1.0 - self.dropout_rate)


class NonLinear(nn.Module):
    def __init__(self, input_size, hidden_size, activation=None):
        super(NonLinear, self).__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.linear = nn.Linear(in_features=input_size, out_features=hidden_size)
        if activation is None:
            self._activate = lambda x: x
        else:
            if not callable(activation):
                raise ValueError("activation must be callable: type={}".format(type(activation)))
            self._activate = activation

        self.reset_parameters()

    def forward(self, x):
        y = self.linear(x)
        return self._activate(y)

    def reset_parameters(self):
        nn.init.xavier_uniform_(self.linear.weight)
        nn.init.zeros_(self.linear.bias)


class Biaffine(nn.Module):
    def __init__(self, in1_features, in2_features, out_features,
                 bias=(True, True)):
        super(Biaffine, self).__init__()
        self.in1_features = in1_features

View on GitHub (pinned to ddb1299bdd)

Solutions

1. Pass a callable: activation=torch.relu (or F.relu, torch.sigmoid); import torch first
2. For identity, pass activation=None
3. If config uses strings, map them: {'relu': torch.relu, 'tanh': torch.tanh}.get(cfg)

Example fix

# before
layer = Scorer(input_size=100, hidden_size=50, activation='relu')
# after
import torch
layer = Scorer(input_size=100, hidden_size=50, activation=torch.relu)

import torch
ACT = {'relu': torch.relu, 'tanh': torch.tanh, 'sigmoid': torch.sigmoid, 'identity': None}
activation = ACT[activation] if isinstance(activation, str) else activation
assert activation is None or callable(activation)

def is_callable_activation(a) -> bool:
    return a is None or callable(a)

When it happens

Trigger: Constructing the SRL span-rank layer (or a config-driven model build) with activation='relu' or activation="tanh" instead of a callable; also passing a class (torch.nn.ReLU) without instantiating when the code expects an instance/function.

Common situations: Config files where activation is a string; porting hyperparameters from Keras/sklearn-style configs into HanLP training scripts.

Understand the failure class

Background: Invalid argument type errors: "must be of type string", "expected X, got Y", and ERR_INVALID_ARG_TYPE explained — this error's family across 15 libraries.

Related errors

• Got average f{average}, expected one of None, 'token', or 'b
• invalid number of tags: {num_tags}
• invalid reduction: {reduction}
• alpha must be float, list of float, or torch.FloatTensor, {}
• Only supports floating point dtypes.

AI-assisted analysis of hankcs/HanLP@ddb1299bdd (2026-08-27). Data as JSON: /api/errors/eb0a0b1829618d8c. Report an issue: GitHub.