Keras自定义梅尔频谱图层构建模型时报'NoneType'不可下标错误

• 自定义MelLayer没有实现compute_output_shape方法，Keras静态构建计算图时无法正确推断层的输出形状，导致后续Conv2D层读取输入通道维度时拿到空值触发报错。
• call方法中全局tf.squeeze(audio)逻辑存在隐患：输入形状为(batch_size, 32000, 1)，全局squeeze会在batch_size为1的场景下错误压缩批次维度，进一步打乱形状推导逻辑。

1. 将call方法中的全局squeeze修改为仅压缩最后一个声道维度，保留批次维度
2. 重写compute_output_shape方法，显式声明层的输出形状，让Keras可以正确做静态形状推断
3. （可选优化）将mel滤波器矩阵的初始化移到build方法中，符合Keras自定义层的规范，避免静态形状推断冲突

修复后的完整MelLayer代码如下：

class MelLayer(tf.keras.layers.Layer):
    def __init__(
        self,
        frame_length=1024,
        frame_step=256,
        fft_length=None,
        sampling_rate=16000,
        num_mel_channels=80,
        freq_min=1,
        freq_max=7600,
        as_3D_tensor=True,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.frame_length = frame_length
        self.frame_step = frame_step
        self.fft_length = fft_length if fft_length else frame_length
        self.sampling_rate = sampling_rate
        self.num_mel_channels = num_mel_channels
        self.freq_min = freq_min
        self.freq_max = freq_max
        self.as_3D_tensor = as_3D_tensor

    def build(self, input_shape):
        # 将mel滤波器初始化移到build方法，避免静态形状冲突
        self.mel_filterbank = tf.signal.linear_to_mel_weight_matrix(
            num_mel_bins=self.num_mel_channels,
            num_spectrogram_bins=self.fft_length // 2 + 1,
            sample_rate=self.sampling_rate,
            lower_edge_hertz=self.freq_min,
            upper_edge_hertz=self.freq_max,
        )
        self.non_trainable_weights.append(self.mel_filterbank)
        super().build(input_shape)

    def call(self, audio, training=True):
        # 仅压缩最后一维声道，保留批次维度
        stft = tf.signal.stft(
            tf.squeeze(audio, axis=-1),
            self.frame_length,
            self.frame_step,
            self.fft_length,
            pad_end=True,
        )

        magnitude = tf.abs(stft)
        mel = tf.matmul(tf.square(magnitude), self.mel_filterbank)
        log_mel_spec = tfio.audio.dbscale(mel, top_db=80)
 
        return tf.expand_dims(log_mel_spec,axis=-1) if self.as_3D_tensor else tf.squeeze(log_mel_spec, axis=-1)

    def compute_output_shape(self, input_shape):
        # 显式声明输出形状
        batch_size = input_shape[0]
        if input_shape[1] is None:
            time_steps = None
        else:
            # pad_end=True时时间步向上取整
            time_steps = (input_shape[1] + self.frame_step - 1) // self.frame_step
        if self.as_3D_tensor:
            return (batch_size, time_steps, self.num_mel_channels, 1)
        else:
            return (batch_size, time_steps, self.num_mel_channels)

    def get_config(self):
        config = super().get_config()
        config.update(
            {
                "frame_length": self.frame_length,
                "frame_step": self.frame_step,
                "fft_length": self.fft_length,
                "sampling_rate": self.sampling_rate,
                "num_mel_channels": self.num_mel_channels,
                "freq_min": self.freq_min,
                "freq_max": self.freq_max,
                "as_3D_tensor": self.as_3D_tensor
            }
        )
        return config

修复后重新构建模型，mel_l的输出形状会被正确推断为(None, 63, 80, 1)，后续Conv2D层可以正常读取通道维度，不会再触发报错。

内容的提问来源于stack exchange，提问作者ABaron