LSTM+VNN实现主动转被动文本训练时维度不匹配错误求助

我用LSTM和VNN实现主动语态到被动语态的文本转换。将分词后的数据输入LSTM，得到形状为(68,1)的特征向量feature_vec，把它作为Vanilla神经网络（VNN）的输入，对应的输出概率矩阵output_prob_matrix形状是(68,10,10)。调用model.fit()时出现以下错误：

Epoch 1/100
Traceback (most recent call last):
 File "/Users/pranavpallavalli/PycharmProjects/pythonProject3/LstmSeq2Seq/ActivePassiveLSTM.py", line 77, in <module>
vnn.fit(feature_vec,output_prob_matrix,32,100)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
raise e.with_traceback(filtered_tb) from None
 File "/var/folders/6q/v9z_sbmd25q2ntjw2pn6hfk80000gn/T/__autograph_generated_fileg6buiumj.py", line 15, in tf__train_function
retval_ = ag__.converted_call(ag__.ld(step_function), (ag__.ld(self), ag__.ld(iterator)), None, fscope)
ValueError: in user code:
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/engine/training.py", line 1051, in train_function  *
    return step_function(self, iterator)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/engine/training.py", line 1040, in step_function  **
    outputs = model.distribute_strategy.run(run_step, args=(data,))
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/engine/training.py", line 1030, in run_step  **
    outputs = model.train_step(data)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/engine/training.py", line 890, in train_step
    loss = self.compute_loss(x, y, y_pred, sample_weight)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/engine/training.py", line 948, in compute_loss
    return self.compiled_loss(
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/engine/compile_utils.py", line 201, in __call__
    loss_value = loss_obj(y_t, y_p, sample_weight=sw)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/losses.py", line 139, in __call__
    losses = call_fn(y_true, y_pred)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/losses.py", line 243, in call  **
    return ag_fn(y_true, y_pred, **self._fn_kwargs)
File "/Users/pranavpallavalli/miniforge3/envs/env_tensorflow/lib/python3.9/site-packages/keras/losses.py", line 1327, in mean_squared_error
    return backend.mean(tf.math.squared_difference(y_pred, y_true), axis=-1)

ValueError: Dimensions must be equal, but are 68 and 10 for '{{node mean_squared_error/SquaredDifference}} = SquaredDifference[T=DT_FLOAT](sequential_1/dense_1/Sigmoid, IteratorGetNext:1)' with input shapes: [68,10], [?,10,10].

模型搭建代码如下：

lstm = Sequential()
lstm.add(LSTM(10,input_shape=(1,10)))
lstm.add(Dense(1))
feature_vec = lstm(lstm_input_sequences)
feature_vec = np.array(feature_vec)
# np.reshape(feature_vec,(68,10,10))
vnn = Sequential()
vnn.add(Input(1,68))
vnn.add(Dense(units=10,activation='sigmoid'))
loss_fn = keras.losses.MeanSquaredError()
vnn.compile(loss='mse',optimizer='adam',metrics=['accuracy'])
print(vnn.summary())
vnn.fit(feature_vec,output_prob_matrix,32,100)

核心问题：输入输出维度不匹配 #

错误日志明确显示，VNN的输出形状是[68,10]，但标签output_prob_matrix的形状是[?,10,10]，两者维度无法对齐，导致MSE损失计算失败。同时VNN的输入层定义也存在错误。

1. 修正VNN输入层 #

原代码中vnn.add(Input(1,68))的写法错误，Keras的Input层需要传入shape参数定义单个样本的特征形状。因为feature_vec是(68,1)，即每个样本有1个特征，所以输入层应改为：

vnn.add(Input(shape=(1,)))

2. 调整VNN输出层，匹配标签形状 #

标签output_prob_matrix是(68,10,10)，意味着每个样本要输出10×10=100个值。需要修改VNN的输出结构，让最终输出形状与标签一致：

vnn = Sequential()
vnn.add(Input(shape=(1,)))
# 可添加中间隐藏层提升模型能力（可选）
vnn.add(Dense(units=128, activation='relu'))
# 先输出100个神经元对应10×10的维度
vnn.add(Dense(units=100, activation='sigmoid'))
# 用Reshape层把输出转换成(10,10)
vnn.add(Reshape((10,10)))

3. 规范fit函数参数 #

原代码中vnn.fit(feature_vec,output_prob_matrix,32,100)的参数顺序容易混淆，建议显式指定参数名：

vnn.fit(feature_vec, output_prob_matrix, batch_size=32, epochs=100)

可选优化：提升LSTM特征表达能力 #

当前LSTM最后只输出1维特征，可能不足以支撑复杂的文本转换任务。可以考虑去掉LSTM后的Dense(1)，直接用LSTM的输出作为VNN输入，比如：

lstm = Sequential()
lstm.add(LSTM(10, input_shape=(1,10)))
# 去掉Dense(1)，得到(68,10)的特征向量
feature_vec = lstm(lstm_input_sequences)
feature_vec = np.array(feature_vec)
# 对应修改VNN输入层为shape=(10,)
vnn.add(Input(shape=(10,)))

这样能保留更多LSTM学到的序列特征，提升模型性能。

内容的提问来源于stack exchange，提问作者abzd