When to use tf.resource and tf.variant?

Feature	tf.resource	tf.variant
Purpose	Handles stateful objects (e.g., variables, tables)	Handles composite or custom types (e.g., ragged tensors, datasets, queues)
Used for	tf.Variable, lookup tables, queues, etc.	Ragged tensors, tf.data.Dataset, distribution values, nested tensors
Stateful?	Yes (tracks mutable state)	Typically stateless or custom serialized
Custom Ops?	Rarely needed	Often used when defining custom ops
User-level use?	Almost never directly	Occasionally when building advanced models/layers
Serialization	Doesn’t serialize to tensors	Can encapsulate structured data in a tensor-like wrapper

Example (handled automatically):

var = tf.Variable(5.0)

print(var.dtype) # tf.float32 (not tf.resource — it’s internal)

Use tf.variant when:

You want to pass complex data structures (like nested lists, ragged arrays) through a graph.

You’re implementing custom operations that need to return or consume non-primitive data types.

Example: RaggedTensor (internally uses tf.variant)

rt = tf.ragged.constant([[1, 2], [3]])

print(rt.dtype) # tf.int32

print(rt.flat_values.dtype) # tf.int32

# Internally wrapped in tf.variant

Example: tf.data pipeline (internally uses tf.variant)

dataset = tf.data.Dataset.range(5)

for x in dataset:

print(x.numpy()) # Output: 0 1 2 3 4

You may work directly with tf.variant:

When working with serialized tensors, like from custom TensorFlow plugins or advanced data types