Bitcasts a tensor from one type to another without copying data.

Given a tensor input, this operation returns a tensor that has the same buffer data as input with datatype type.

If the input datatype T is larger than the output datatype type then the shape changes from [...] to [..., sizeof(T)/sizeof(type)].

If T is smaller than type, the operator requires that the rightmost dimension be equal to sizeof(type)/sizeof(T). The shape then goes from [..., sizeof(type)/sizeof(T)] to [...].

tf.bitcast() and tf.cast() work differently when real dtype is casted as a complex dtype (e.g. tf.complex64 or tf.complex128) as tf.cast() make imaginary part 0 while tf.bitcast() gives module error. For example,

Example 1:

a = [1., 2., 3.]
equality_bitcast = tf.bitcast(a, tf.complex128)
Traceback (most recent call last):

InvalidArgumentError: Cannot bitcast from 1 to 18 [Op:Bitcast]
equality_cast = tf.cast(a, tf.complex128)
tf.Tensor([1.+0.j 2.+0.j 3.+0.j], shape=(3,), dtype=complex128)

Example 2:

tf.bitcast(tf.constant(0xffffffff, dtype=tf.uint32), tf.uint8)
<tf.Tensor: shape=(4,), dtype=uint8, numpy=array([255, 255, 255, 255], dtype=uint8)>

Example 3:

x = [1., 2., 3.]
y = [0., 2., 3.]
equality= tf.equal(x,y)
equality_cast = tf.cast(equality,tf.float32)
equality_bitcast = tf.bitcast(equality_cast,tf.uint8)
tf.Tensor([False True True], shape=(3,), dtype=bool)