What causes "NotImplementedError" in d2l-ai/d2l-zh?

Subclass and override call: class MyEncoder(d2l.Encoder): def call(self, X, *args, **kwargs): ...

d2l-ai/d2l-zh · error · NotImplementedError

NotImplementedError

Error message

NotImplementedError

What it means

A NotImplementedError raised by d2l.tensorflow.Encoder.call. Encoder is the abstract interface of the encoder-decoder architecture (sec_encoder-decoder): it is a tf.keras.layers.Layer whose call is intentionally left unimplemented. Subclasses (Seq2SeqEncoder, TransformerEncoder, ...) must override call; instantiating the base class and invoking it as a layer triggers the error.

Source

Thrown at d2l/tensorflow.py:881

    text = preprocess_nmt(read_data_nmt())
    source, target = tokenize_nmt(text, num_examples)
    src_vocab = d2l.Vocab(source, min_freq=2,
                          reserved_tokens=['<pad>', '<bos>', '<eos>'])
    tgt_vocab = d2l.Vocab(target, min_freq=2,
                          reserved_tokens=['<pad>', '<bos>', '<eos>'])
    src_array, src_valid_len = build_array_nmt(source, src_vocab, num_steps)
    tgt_array, tgt_valid_len = build_array_nmt(target, tgt_vocab, num_steps)
    data_arrays = (src_array, src_valid_len, tgt_array, tgt_valid_len)
    data_iter = d2l.load_array(data_arrays, batch_size)
    return data_iter, src_vocab, tgt_vocab

class Encoder(tf.keras.layers.Layer):
    """编码器-解码器架构的基本编码器接口"""
    def __init__(self, **kwargs):
        super(Encoder, self).__init__(**kwargs)

    def call(self, X, *args, **kwargs):
        raise NotImplementedError

class Decoder(tf.keras.layers.Layer):
    """编码器-解码器架构的基本解码器接口

    Defined in :numref:`sec_encoder-decoder`"""
    def __init__(self, **kwargs):
        super(Decoder, self).__init__(**kwargs)

    def init_state(self, enc_outputs, *args):
        raise NotImplementedError

    def call(self, X, state, **kwargs):
        raise NotImplementedError

class EncoderDecoder(tf.keras.Model):
    """编码器-解码器架构的基类

    Defined in :numref:`sec_encoder-decoder`"""

View on GitHub (pinned to e6b18ccea7)

Solutions

Subclass and override call: class MyEncoder(d2l.Encoder): def call(self, X, *args, **kwargs): ...
If porting from the PyTorch edition, rename forward -> call in TensorFlow subclasses.
Use the provided concrete implementations, e.g. d2l.Seq2SeqEncoder(vocab_size, num_hiddens, num_layers, dropout), instead of the abstract base.
Check for typos in the method name (call vs calls) and correct signature (self, X, *args, **kwargs).

Example fix

# before
enc = d2l.Encoder()
enc(X)  # NotImplementedError
# after
class MyEncoder(d2l.Encoder):
    def call(self, X, *args, **kwargs):
        return tf.identity(X)
enc = MyEncoder()
enc(X)

Defensive patterns

Strategy: type-guard

Type guard

def is_concrete_encoder(enc) -> bool:
    return (isinstance(enc, d2l.Encoder)
            and type(enc).call is not d2l.Encoder.call)

Try / catch

try:
    enc(X)
except NotImplementedError:
    raise TypeError(f'{type(enc).__name__} must override Encoder.call(X, *args, **kwargs)') from None

Prevention

In TensorFlow subclasses, the method must be named call (not forward).
Prefer library encoders (d2l.Seq2SeqEncoder) unless writing a custom architecture.
Add a smoke test that runs one forward pass on a tiny batch right after constructing any new encoder.

When it happens

Trigger: Instantiating d2l.Encoder() directly and calling encoder(X); a custom encoder subclass whose forward method is misspelled (e.g. 'forward' instead of 'call' in the TF API) so the base call runs; type-testing code that calls call on an arbitrary Encoder instance.

Common situations: Porting PyTorch d2l code to TensorFlow: writing def forward(self, X) on a TF subclass silently falls through to the base call; omitting the override entirely while developing a new architecture chapter; IDE auto-generating __init__ but not call.

Related errors

AI-assisted analysis of d2l-ai/d2l-zh@e6b18ccea7 (2026-08-14). Data as JSON: /api/errors/26cf03749cf2c26b. Report an issue: GitHub.