[FIXED] Best way to convert Tensor from [24, 512, 768, 1] to [24, 512, 14, 14]

Issue

I had an incompatibility issue with PyTorch tensor shape. Hence, I need to convert one tensor from shape [24, 512, 768, 1] to [24, 512, 14, 14]. What is the best way while trying to preserve as much info as possible about the original tensor representation?

Solution

I think this is an optimal solution, regardless of the data preservation you want to include:

import tensorflow as tf

t = tf.constant([[[[1]]*768]*512]*24)
t = tf.reshape(tf.constant(t.numpy()[:, :, :196]), (24, 512, 14, 14))
print(t.shape)

Output:

(24, 512, 14, 14)

Answered By - Cardstdani