如何将Keras生成的Tensor转为TensorType？或求适配Tensor的方案

Hey there! Let's break down this Tensor type mismatch issue you're facing and get it fixed.

The error you're seeing happens because the code you're passing your Keras Tensor into expects a TensorType (likely an older tensor type from frameworks like Theano, or a custom-defined type), but you're providing a modern TensorFlow tf.Tensor (which is what Keras uses under the hood with the TF backend). Mixing these two incompatible types triggers the conversion failure.

Since Keras now tightly integrates with TensorFlow, it's much cleaner and more reliable to adjust your code to operate entirely with tf.Tensor objects instead of forcing type conversions. Here's how to do it step by step:

1. Convert your self.pi ndarray to a tf.Tensor
   First, turn your numpy array into a TensorFlow tensor so it matches the type of your inputs variable:

   import tensorflow as tf
   
   # Convert self.pi from ndarray to tf.Tensor (match dtype to your inputs if needed)
   self.pi_tensor = tf.convert_to_tensor(self.pi, dtype=tf.float32)
   

2. Replace numpy operations with TensorFlow equivalents
   If your code uses numpy functions (like np.mean, np.sum, etc.) on self.pi, swap them out for TensorFlow's corresponding functions to keep everything within the computation graph. For example:

   • Replace np.mean(self.pi) with tf.reduce_mean(self.pi_tensor)
   • Replace np.sum(self.pi, axis=0) with tf.reduce_sum(self.pi_tensor, axis=0)

3. Update the target code to accept tf.Tensor inputs
   Ensure the section of code that takes inputs is set up to handle tf.Tensor objects. Since Keras Tensors are just tf.Tensors with extra Keras-specific attributes, this should work seamlessly once all variables are in TensorFlow's tensor format.

If you absolutely can't modify the target code to use tf.Tensor (e.g., it's a legacy library dependent on Theano), you can try converting your Keras Tensor to a Theano TensorType. Note this requires having Theano installed and set as Keras' backend, which is not common in modern workflows:

from keras import backend as K
import theano.tensor as T

# First extract the numpy array from the Keras Tensor
inputs_np = K.eval(inputs)
# Convert to Theano's TensorType
inputs_tensor_type = T.as_tensor_variable(inputs_np)

Warning: This approach breaks graph-based operations (like training loops) because you're pulling the tensor out of the TensorFlow computation graph. Only use this if you have no other option.

内容的提问来源于stack exchange，提问作者Alk