is there a way to transpose a tensor without using the transpose function in tensorflow?

Question

I have a model that i want to port to tflite micro. However, when i run the code it gives me the following error:

Didn't find op for builtin opcode 'TRANSPOSE' version '2'
Failed to get registration from op code d

I assume that the transpose function is not supported in tflite micro. i also tried replacing it with a PERMUTE layer but it seems that it uses tf.transpose under the hood. Here is the part of my model where i try to traspose:

output = tf.reshape(output, (img_width // B, B, img_height // B, B), name="reshape_in")
output = Permute((2, 1, 3), name="transpose_in")(output)

is there any other way i could perform this transpose without calling tf.transpose? Maybe using reshape?

Answer 1

I had a similar issue when trying to run a model with TFLite's GPU delegate, which does not support the transpose operation.

One possibility is to use a combination of strided slices, tf.reshape and concat operations:

shape = (3,4,5)
a = tf.random.uniform(shape)
a_t = tf.transpose(a,(1,0,2)) # permuting first and second axis
a_concat = tf.concat([tf.reshape(a[i:i+1,:,:],(shape[1],1,shape[2])) for i in range(shape[0])],axis=1)
tf.debugging.assert_equal(a_t,a_concat)

Notes:

• The shape of the tensor needs to be known in advance.
• it uses tf.concat because PACK/tf.stack is not available on the GPU delegate
• it uses a[i:i+1,:] instead of a[i] because strided slices that remove an axis are not supported on the GPU delegate
• Performances are likely to be bad, especially if the dimension looped through is big: what this trick does is essentially unrolling the operation on one dimension, creating as many reshape node in the graph as the size of the dimension.