TensorFlow自定义Encoder/Decoder模型无法保存问题求助

问题根源 #

你的自定义Encoder和Decoder模型无法保存的核心问题在于**get_config方法实现错误**：

• Decoder类中super().get_config(self)调用错误，正确写法是不带self参数的super().get_config()
• get_config直接返回了层对象（比如self.lstm1），而TensorFlow无法序列化层实例，必须返回可序列化的层配置字典
• 模型初始化依赖的外部参数（h_size、vocab_len、emb_size）未纳入配置，导致重建模型时无法还原结构

解决方案1：正确实现get_config和from_config #

修正后的模型代码确保配置可序列化且能正确重建模型：

修正后的Encoder类 #

class Encoder(tf.keras.Model):
    def __init__(self, h_size):
        super().__init__()
        self.h_size = h_size
        self.lstm1 = tf.keras.layers.LSTM(h_size * 2, return_sequences=False, return_state=True, name='lstm1_enc')
        self.lstm2 = tf.keras.layers.LSTM(h_size * 2, return_sequences=True, return_state=True, name='lstm2_enc')
        self.lstm3 = tf.keras.layers.LSTM(h_size * 4, return_sequences=True, return_state=True, name='lstm3_enc')
        
    def call(self, x):
        state = []
        out, h, c = self.lstm3(x)
        state.append((h, c))
        out, h, c = self.lstm2(out)
        state.append((h, c))
        out, h, c = self.lstm1(out)
        state.append((h, c))
        
        return state
    
    def get_config(self):
        config = super().get_config()
        config.update({
            'h_size': self.h_size,
            'lstm1': self.lstm1.get_config(),
            'lstm2': self.lstm2.get_config(),
            'lstm3': self.lstm3.get_config()
        })
        return config
    
    @classmethod
    def from_config(cls, config):
        model = cls(config['h_size'])
        model.lstm1 = tf.keras.layers.LSTM.from_config(config['lstm1'])
        model.lstm2 = tf.keras.layers.LSTM.from_config(config['lstm2'])
        model.lstm3 = tf.keras.layers.LSTM.from_config(config['lstm3'])
        return model

修正后的Decoder类 #

class Decoder(tf.keras.Model):
    def __init__(self, vocab_len, emb_size, h_size):
        super().__init__()
        self.vocab_len = vocab_len
        self.emb_size = emb_size
        self.h_size = h_size
        self.embed = tf.keras.layers.Embedding(vocab_len, emb_size, name='embed')
        self.lstm1 = tf.keras.layers.LSTM(h_size * 2, return_sequences=True, return_state=True, name='lstm1_dec')
        self.lstm2 = tf.keras.layers.LSTM(h_size * 2, return_sequences=True, return_state=True, name='lstm2_dec')
        self.lstm3 = tf.keras.layers.LSTM(h_size * 4, return_sequences=True, return_state=True, name='lstm3_dec')
        self.fc = tf.keras.layers.Dense(vocab_len, activation='softmax', name='fc')
        
    def call(self, x, init_state):
        state = []
        out = self.embed(x)
        out, h, c = self.lstm1(out, initial_state=init_state[2])
        state.append((h, c))
        out, h, c = self.lstm2(out, initial_state=init_state[1])
        state.append((h, c))
        out, h, c = self.lstm3(out, initial_state=init_state[0])
        state.append((h, c))
        out = self.fc(out)
        
        return out, state

    def get_config(self):
        config = super().get_config()
        config.update({
            'vocab_len': self.vocab_len,
            'emb_size': self.emb_size,
            'h_size': self.h_size,
            'embed': self.embed.get_config(),
            'lstm1': self.lstm1.get_config(),
            'lstm2': self.lstm2.get_config(),
            'lstm3': self.lstm3.get_config(),
            'fc': self.fc.get_config()
        })
        return config
    
    @classmethod
    def from_config(cls, config):
        model = cls(config['vocab_len'], config['emb_size'], config['h_size'])
        model.embed = tf.keras.layers.Embedding.from_config(config['embed'])
        model.lstm1 = tf.keras.layers.LSTM.from_config(config['lstm1'])
        model.lstm2 = tf.keras.layers.LSTM.from_config(config['lstm2'])
        model.lstm3 = tf.keras.layers.LSTM.from_config(config['lstm3'])
        model.fc = tf.keras.layers.Dense.from_config(config['fc'])
        return model

初始化与保存 #

传入实际参数初始化模型，之后即可正常保存：

# 替换为你实际的参数值
h_size = 64
vocab_len = 1000
emb_size = 128
random_vector_len = 32

encoder_model = Encoder(h_size)
decoder_model = Decoder(vocab_len, emb_size, h_size)

# 保持原seq2seq构建代码不变
encoder_inputs = tf.keras.layers.Input(shape=(None, random_vector_len))
decoder_inputs = tf.keras.layers.Input(shape=(None, ))

enc_state = encoder_model(encoder_inputs)
decoder_outputs, _ = decoder_model(decoder_inputs, enc_state)

seq2seq = tf.keras.Model([encoder_inputs, decoder_inputs], decoder_outputs)

# 执行保存
tf.keras.models.save_model(encoder_model, 'encoder1', overwrite=True)
tf.keras.models.save_model(decoder_model, 'decoder', overwrite=True)

解决方案2：仅保存权重（快速替代方案） #

若不想修改get_config，可使用save_weights仅保存权重，后续重建模型结构后加载：

保存权重 #

encoder_model.save_weights('encoder_weights.h5')
decoder_model.save_weights('decoder_weights.h5')

加载权重 #

# 重建Encoder并加载权重
encoder_model = Encoder(h_size)
dummy_enc_input = tf.random.normal((1, 10, random_vector_len))
encoder_model(dummy_enc_input)
encoder_model.load_weights('encoder_weights.h5')

# 重建Decoder并加载权重
decoder_model = Decoder(vocab_len, emb_size, h_size)
dummy_dec_input = tf.random.uniform((1, 5), maxval=vocab_len, dtype=tf.int32)
dummy_state = [
    (tf.random.normal((1, h_size*4)), tf.random.normal((1, h_size*4))),
    (tf.random.normal((1, h_size*2)), tf.random.normal((1, h_size*2))),
    (tf.random.normal((1, h_size*2)), tf.random.normal((1, h_size*2)))
]
decoder_model(dummy_dec_input, dummy_state)
decoder_model.load_weights('decoder_weights.h5')

内容的提问来源于stack exchange，提问作者crn4t