Numpy Indexing Operations with Tensorflow

Question

I want to update the values of a tensor like:

ldream[w,:,x,y,z] = 0

But keep getting an error saying:

TypeError: 'Tensor' object does not support item assignment

Answer 1

It seems ldream is a tensorflow variable, which unfortunately cannot be directly assigned values the same way as numpy arrays.

To update the value of a tensor in tensorflow, you can make an assignment operation and then run (eval) that operation. Here is an example on how to do this:

Tensorflow: How to modify the value in tensor

The reason for this is that when you code your tf variables and operations, you are actually "staging" them to happen at a later time (such as when you do sess.run ); like a blueprint of how tensorflow will actually run.

Answer 2

Fixed this by creating a boolean mask using a numpy array:

ldream_mask = np.zeros(ldream.shape, dtype=np.bool)

Then selecting the desired indices and marking them as True

ldream_mask[w,:,x,y,z] =  True

Then using tf.where to update the desired indices using ldream_mask