Keras有状态LSTM错误:从shape[32,1]的张量中指定了shape[4,1]的列表

Question

使用此代码，我在尝试使用X_test运行预测时出错。该错误发生在拟合之后，同时执行y_pred = model.predict(X_test)。

X_train.input_shape()
-> (784, 300, 7)
y_train.input_shape()
-> (784, 300, 1)
X_test.input_shape()
-> (124,300,7)
y_test.input_shape()
-> (124,300,1)

batchsize = 4
model = Sequential()
model.add(Masking(mask_value=0, batch_input_shape=(batchsize, len(X_train[0]),len(X_train[0][0]))))
model.add(Bidirectional(LSTM(200, return_sequences=True, stateful=True)))
model.add(TimeDistributed(Dense(len(classdict))))
model.compile(loss="sparse_categorical_crossentropy",
optimizer=Adam(0.001),
metrics=['accuracy'])
model.summary()
model.fit(X_train, y_train, epochs=1, batch_size=batchsize)
y_pred = model.predict(X_test) 

错误：

Traceback (most recent call last):
File "lib/python3.7/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
raise e.with_traceback(filtered_tb) from None
File "lib/python3.7/site-packages/tensorflow/python/eager/execute.py", line 55, in quick_execute
inputs, attrs, num_outputs)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Graph execution error:
Specified a list with shape [4,1] from a tensor with shape [32,1]
[[{{node TensorArrayUnstack_1/TensorListFromTensor}}]]
[[sequential/bidirectional/backward_lstm/PartitionedCall]] [Op:__inference_predict_function_13360]

在我看来，错误可能与所选的批量大小有关。在我的研究中，我读到有状态LSTM的样本数量必须被批量大小整除。所以我搜索X_train和X_test的样本数的最大公约数，所以GCD(124784(=4=批大小。然而，现在出现了这个错误，我已经尝试了不同的批处理大小，但随后出现了其他错误。有人对此有想法/解决办法吗？

Answer 1

您只需要确保训练样本的数量可以平均除以批次大小。下面是一个工作示例：

import tensorflow as tf
X_train = tf.random.normal((784, 300, 7))
y_train = tf.random.normal((784, 300, 1))
X_test = tf.random.normal((124,300,7))
batchsize = 4
model = tf.keras.Sequential()
model.add(tf.keras.layers.Masking(mask_value=0, batch_input_shape=(batchsize, len(X_train[0]),len(X_train[0][0]))))
model.add(tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(200, return_sequences=True, stateful=True)))
model.add(tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(1)))
model.compile(loss="mse",
optimizer=tf.keras.optimizers.Adam(0.001))
model.summary()
model.fit(X_train, y_train, batch_size=batchsize, epochs=1)

Model: "sequential_5"
_________________________________________________________________
Layer (type)                Output Shape              Param #   
=================================================================
masking_5 (Masking)         (4, 300, 7)               0         

bidirectional_5 (Bidirectio  (4, 300, 400)            332800    
nal)                                                            

time_distributed_5 (TimeDis  (4, 300, 1)              401       
tributed)                                                       

=================================================================
Total params: 333,201
Trainable params: 333,201
Non-trainable params: 0
_________________________________________________________________
196/196 [==============================] - 47s 131ms/step - loss: 1.0052
<keras.callbacks.History at 0x7fee901f8950>

并且在预测时明确设置批次大小，因为根据文档，如果未指定，batch_size将默认为32:

X_test = tf.random.normal((124,300,7))
y_pred = model.predict(X_test, batch_size=batchsize) 
print(y_pred.shape)

(124, 300, 1)

还要确保你的批次在任何地方都是一样的。