lets say I have a 64x64x64 3D image. I also have a vector x which is length 64.
I want to take the 'argmax(x)' layer like so:
2d_image = 3d_image[:,argmax(x),:]
more precise (for tensorsflow):
def extract_slice(x,3d_image):
slice_index = tf.math.argmax(x,axis=1,output_type=tf.dtype.int32) #it has to be int for indexing
return 3d_image[:,slice_index,:]
The error is:
Only integers, slices (`:`), ellipsis (`...`), tf.newaxis (`None`) and scalar tf.int32/tf.int64 tensors are valid indices, got <tf.Tensor 'ArgMax_50:0' shape=(None,) dtype=int32>
The np.shape of the paramters are:
3d_image shape is (None, 64, 64, 64, 1)
x shape is (None, 64)
slice_index shape is (None,)
->the ,1 dimension at 3d_image shape is because that its a sample from array.. I don't think its matter
I know that None shape is the batch size, which is unknown, but the others looks excellent.. so what is the problem?
From what I understand, it looks like the index is not int32, but I actually DID cast it to tf.int so what could be the problem?? maybe int32 is differ from tf.int32? or maybe the index method I used is not valid in tensorflow? maybe it should be a function something like that: tf.index(image,[:,slice_index,:])..?
Thanks!
Argmax returns 1D tensor. Convert it to scalar:
slice_index = tf.reshape(slice_index, ())