传递给参数“indices”的值的数据类型 float32 不在允许值列表中：uint8、int8、int32、int64

Question

我正在尝试为我的手写文本识别模型构建一个用于损失计算的 CTC 层，但遇到这些错误

TypeError：调用CTC_Layer.call()时遇到异常。

传递给参数“indices”的值的数据类型 float32 不在允许值列表中：uint8、int8、int32、int64

class CTC_Layer(Layer):
    def __init__(self, name=None):
        super(CTC_Layer, self).__init__(name='ctc_loss')
        self.loss_fn = tensorflow.nn.ctc_loss
        
    def call(self, y_true, y_pred):
        batch_length = tf.cast(tf.shape(y_true)[0], "int64")
        input_length = tf.cast(tf.shape(y_pred)[1], "int64")
        label_length = tf.cast(tf.shape(y_true)[1], "int64")
        
        input_length = input_length * tf.ones(shape=(batch_length,), dtype="int64")
        label_length = label_length * tf.ones(shape=(batch_length,), dtype="int64")
        
        loss = self.loss_fn(y_true, y_pred, input_length, label_length)
        self.add_loss(loss)
        
        return y_pred

手写文本识别模型由 2-D CNN、双向 LSTM 组成

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[133], line 1
----> 1 model = build_model()
      2 model.summary()

Cell In[132], line 31, in build_model()
     24 x = layers.Bidirectional(
     25     layers.LSTM(64, return_sequences=True, dropout=0.25)
     26 )(x)
     28 x = layers.Dense(
     29     len((char_to_num.get_vocabulary()))+2, activation='softmax', name='dense2'
     30 )(x)
---> 31 outputs = CTC_Layer(name="ctc_loss")(labels, x)
     34 model = keras.models.Model(inputs=[input_image, labels], outputs=outputs)
     36 model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001))

File /opt/conda/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py:122, in filter_traceback.<locals>.error_handler(*args, **kwargs)
    119     filtered_tb = _process_traceback_frames(e.__traceback__)
    120     # To get the full stack trace, call:
    121     # `keras.config.disable_traceback_filtering()`
--> 122     raise e.with_traceback(filtered_tb) from None
    123 finally:
    124     del filtered_tb

Cell In[93], line 15, in CTC_Layer.call(self, y_true, y_pred)
     12 input_length = input_length * tf.ones(shape=(batch_length,), dtype="int64")
     13 label_length = label_length * tf.ones(shape=(batch_length,), dtype="int64")
---> 15 loss = self.loss_fn(y_true, y_pred, input_length, label_length)
     16 self.add_loss(loss)
     18 # At test time, just return the computed predictions.

TypeError: Exception encountered when calling CTC_Layer.call().

Value passed to parameter 'indices' has DataType float32 not in list of allowed values: uint8, int8, int32, int64

Arguments received by CTC_Layer.call():
  • args=('<KerasTensor shape=(None, None), dtype=float32, sparse=None, name=label>', '<KerasTensor shape=(None, 32, 79), dtype=float32, sparse=False, name=keras_tensor_317>')
  • kwargs=<class 'inspect._empty'>]

有人知道如何解决这个问题吗？预先感谢！

Answer 1

在您的

CTC_Layer.call

函数中，在将标签输入损失函数之前，将它们转换为适当的整数类型。这是更新：

def call(self, y_true, y_pred):
    # ... (rest of your code)

    # Cast labels to int32 (choose appropriate type if needed)
    y_true = tf.cast(y_true, dtype="int32")

    # ... (rest of your code)