新版本TensorFlow中使用legacy_seq2seq.embedding_rnn_seq2seq报错求助

问题原因 #

tf.legacy_seq2seq是TensorFlow 1.x中的旧版序列建模模块，在TensorFlow 2.x中已被移除，仅在tf.compat.v1兼容模块中保留部分接口。你的代码中混用了直接调用tf.legacy_seq2seq和tf.compat.v1.legacy_seq2seq的写法，导致出现属性不存在的报错。

解决方案 #

方案一：兼容TensorFlow 1.x的快速修复 #

将所有tf.legacy_seq2seq的调用统一替换为tf.compat.v1.legacy_seq2seq，同时启用TF1兼容模式：

1. 修改embedding_rnn_seq2seq调用行：

decoderOutputs, states = tf.compat.v1.legacy_seq2seq.embedding_rnn_seq2seq(
    self.encoderInputs,  # List<[batch=?, inputDim=1]>, list of size args.maxLength
    self.decoderInputs,  # For training, we force the correct output (feed_previous=False)
    encoDecoCell,
    self.textData.getVocabularySize(),
    self.textData.getVocabularySize(),  # Both encoder and decoder have the same number of class
    embedding_size=self.args.embeddingSize,  # Dimension of each word
    output_projection=outputProjection.getWeights() if outputProjection else None,
    feed_previous=bool(self.args.test)
    # When we test (self.args.test), we use previous output as next input (feed_previous)
)

2. 在代码开头添加兼容模式启用代码：

import tensorflow as tf
tf.compat.v1.disable_eager_execution()

方案二：迁移到TensorFlow 2.x原生API（长期推荐） #

TF2.x提供了更灵活的Keras接口实现seq2seq模型，替代旧版legacy_seq2seq：

1. 重构编码器-解码器结构
   用Embedding层处理词嵌入，LSTM/GRU作为循环单元，手动构建训练和推理模型：

# 编码器
encoder_embedding = tf.keras.layers.Embedding(
    input_dim=self.textData.getVocabularySize(),
    output_dim=self.args.embeddingSize
)
encoder_lstm = tf.keras.layers.LSTM(encoDecoCell.output_size, return_state=True)

encoder_inputs = tf.keras.Input(shape=(self.args.maxLengthEnco,))
x = encoder_embedding(encoder_inputs)
_, encoder_h, encoder_c = encoder_lstm(x)
encoder_states = [encoder_h, encoder_c]

# 训练解码器
decoder_embedding = tf.keras.layers.Embedding(
    input_dim=self.textData.getVocabularySize(),
    output_dim=self.args.embeddingSize
)
decoder_lstm = tf.keras.layers.LSTM(encoDecoCell.output_size, return_sequences=True, return_state=True)
decoder_dense = tf.keras.layers.Dense(self.textData.getVocabularySize(), activation='softmax')

decoder_inputs = tf.keras.Input(shape=(self.args.maxLengthDeco,))
x = decoder_embedding(decoder_inputs)
decoder_outputs, _, _ = decoder_lstm(x, initial_state=encoder_states)
decoder_outputs = decoder_dense(decoder_outputs)

# 训练模型定义
train_model = tf.keras.Model([encoder_inputs, decoder_inputs], decoder_outputs)

# 推理解码器（自回归模式）
decoder_state_input_h = tf.keras.Input(shape=(encoDecoCell.output_size,))
decoder_state_input_c = tf.keras.Input(shape=(encoDecoCell.output_size,))
decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]

x = decoder_embedding(decoder_inputs)
decoder_outputs, dec_h, dec_c = decoder_lstm(x, initial_state=decoder_states_inputs)
decoder_states = [dec_h, dec_c]
decoder_outputs = decoder_dense(decoder_outputs)

inference_model = tf.keras.Model(
    [decoder_inputs] + decoder_states_inputs,
    [decoder_outputs] + decoder_states
)

2. 替换损失函数
   用TF2原生损失函数结合掩码处理变长序列：

loss_obj = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False, reduction='none')

def compute_loss(real, pred):
    mask = tf.math.logical_not(tf.math.equal(real, 0))
    loss = loss_obj(real, pred)
    mask = tf.cast(mask, dtype=loss.dtype)
    loss *= mask
    return tf.reduce_mean(loss)

3. 重构训练流程
   使用TF2的优化器和自定义训练循环：

optimizer = tf.keras.optimizers.Adam(
    learning_rate=self.args.learningRate,
    beta_1=0.9,
    beta_2=0.999,
    epsilon=1e-08
)

@tf.function
def train_step(enc_inputs, dec_inputs, dec_targets):
    with tf.GradientTape() as tape:
        preds = train_model([enc_inputs, dec_inputs], training=True)
        loss = compute_loss(dec_targets, preds)
    grads = tape.gradient(loss, train_model.trainable_variables)
    optimizer.apply_gradients(zip(grads, train_model.trainable_variables))
    return loss

注意事项 #

• 方案一适合快速修复旧代码，建议搭配TensorFlow 1.15版本使用，避免兼容性问题；
• 方案二需要重构较多代码，但符合TF2设计理念，支持即时执行和分布式训练，长期维护更便捷。

内容的提问来源于stack exchange，提问作者Kritharth Singh