InvalidArgumentError：无法计算 Sub，因为输入 #1（从零开始）预计是 uint8 张量，但实际上是浮点张量 [Op:Sub]

Question

问题

请帮助了解错误原因以及如何解决。

代码

import tensorflow as tf
import numpy as np

fashion_mnist = tf.keras.datasets.fashion_mnist
(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()
x_full = np.concatenate((x_train, x_test), axis=0)

layer = tf.keras.layers.experimental.preprocessing.Normalization()
layer.adapt(x_full)
layer(x_train)

错误

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
<ipython-input-16-699c47b6db55> in <module>
----> 1 ds = layer(x_train)

~/conda/envs/tensorflow/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, *args, **kwargs)
    966           with base_layer_utils.autocast_context_manager(
    967               self._compute_dtype):
--> 968             outputs = self.call(cast_inputs, *args, **kwargs)
    969           self._handle_activity_regularization(inputs, outputs)
    970           self._set_mask_metadata(inputs, outputs, input_masks)

~/conda/envs/tensorflow/lib/python3.7/site-packages/tensorflow/python/keras/layers/preprocessing/normalization.py in call(self, inputs)
    109     mean = array_ops.reshape(self.mean, self._broadcast_shape)
    110     variance = array_ops.reshape(self.variance, self._broadcast_shape)
--> 111     return (inputs - mean) / math_ops.sqrt(variance)
    112 
    113   def compute_output_shape(self, input_shape):

~/conda/envs/tensorflow/lib/python3.7/site-packages/tensorflow/python/ops/math_ops.py in binary_op_wrapper(x, y)
    982     with ops.name_scope(None, op_name, [x, y]) as name:
    983       if isinstance(x, ops.Tensor) and isinstance(y, ops.Tensor):
--> 984         return func(x, y, name=name)
    985       elif not isinstance(y, sparse_tensor.SparseTensor):
    986         try:

~/conda/envs/tensorflow/lib/python3.7/site-packages/tensorflow/python/ops/gen_math_ops.py in sub(x, y, name)
  10098         pass  # Add nodes to the TensorFlow graph.
  10099     except _core._NotOkStatusException as e:
> 10100       _ops.raise_from_not_ok_status(e, name)
  10101   # Add nodes to the TensorFlow graph.
  10102   _, _, _op, _outputs = _op_def_library._apply_op_helper(

~/conda/envs/tensorflow/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in raise_from_not_ok_status(e, name)
   6651   message = e.message + (" name: " + name if name is not None else "")
   6652   # pylint: disable=protected-access
-> 6653   six.raise_from(core._status_to_exception(e.code, message), None)
   6654   # pylint: enable=protected-access
   6655 

~/conda/envs/tensorflow/lib/python3.7/site-packages/six.py in raise_from(value, from_value)

InvalidArgumentError: cannot compute Sub as input #1(zero-based) was expected to be a uint8 tensor but is a float tensor [Op:Sub]

尝试

尝试了 dtype arg 但同样的错误。

layer = tf.keras.layers.experimental.preprocessing.Normalization(dtype='float32')

Divide by 1.0 fixed the issue but not sure the original cause.

x_full = np.concatenate((x_train, x_test), axis=0) / 1.0
x_train = x_train / 1.0

Does Keras only works with float32?

Related issues

• What does InvalidArgumentError in tensorflow 2 mean?
• InvalidArgumentError: cannot compute MatMul as input #0(zero-based) was expected to be a float tensor but is a double tensor [Op:MatMul]

Answer 1

原因是preprocessing.Normalization预期的float32，但你的数据是uint8这样的，因此出现了错误。

这实际上是 Tensorflow 的问题，而不是 Keras 本身的问题，因为计算速度更快。

提醒：float 和 int 计算在处理器中的不同位置，并且每个处理器对不同数据类型的性能不同，例如 nvidia 的 GPU 比 16 更快，float32而float16ARM CPU 的 16 更快。

Pytorch 也需要两个变量具有相同的数据类型，否则它将无法工作。

在Python中用一个整数除以一个浮点数会自动给你一个新的浮点数，x_train = x_train / 1.0将得到x_train float32（或者float64或者float16取决于你在这里有什么~/.keras/keras.json）float32。