Normalize operation inputs
A custom Python operation receives lists, scalars, or NumPy arrays and needs a consistent Tensor input.
TensorFlow · Function
tf.convert_to_tensor
Converts supported Python and array-like values into a Tensor. The element type is inferred unless dtype is supplied, while dtype_hint provides a fallback preference.
How it works
The function uses TensorFlow's registered conversion for the supplied value and returns a Tensor. It infers the element type unless dtype is provided; dtype_hint supplies a fallback preference only when dtype is omitted and conversion to that type is possible.
Syntax
tf.convert_to_tensor(value, dtype=None, dtype_hint=None, name=None) -> tf.TensorParameters
| Name | Type | Description |
|---|---|---|
value | Any | An object with a registered Tensor conversion function. |
dtype | tf.DType | Optional element type for the returned Tensor; if omitted, it is inferred from value. |
dtype_hint | tf.DType | Optional element type used as a soft preference when dtype is None. |
name | string | Optional name used if a new Tensor is created. |
Returns
A Tensor based on value.
Key points
- The function turns a supported tensor-like value into a Tensor.
- Common accepted inputs include existing Tensors, NumPy arrays, and Python lists.
- Python scalar values are also valid inputs.
- Without dtype, TensorFlow determines the element type from the input.
- name labels a newly created Tensor when a name is supplied.
- TensorFlow operation constructors use this conversion for tensor-valued inputs, enabling tensor-like arguments.
Common mistakes
- dtype_hint is only a soft preference: TensorFlow ignores it when conversion to that type cannot be performed.
- A Tensor with a different dtype can raise ValueError in graph mode when an explicit dtype is requested.
- Values without a registered conversion function cause TypeError.
Example: Convert a list to a Tensor
import tensorflow as tf
values = tf.convert_to_tensor([1, 2, 3], dtype=tf.float32)
print(values)
Converts a Python list into a float32 Tensor that can be passed to TensorFlow operations.
When to use it
Set the input dtype
Data must enter a TensorFlow computation with a specified element type.
Support tensor-like arguments
An operation should accept tensor-like values without requiring callers to construct Tensors first.
Interview questions
QWhat does tf.convert_to_tensor do?
It converts supported Python values, NumPy arrays, lists, scalars, and existing Tensor objects into a Tensor. TensorFlow operations use this behavior so many operations can accept tensor-like inputs.
Follow-up: What happens when the input type is unsupported? TensorFlow raises TypeError if no conversion function is registered for the value and requested dtype.
QHow are dtype and dtype_hint different?
dtype explicitly specifies the returned tensor's element type. If dtype is absent, TensorFlow infers the type from value.
Follow-up: How does dtype_hint differ? It is used only when dtype is None and acts as a soft preference rather than a required type.
QWhat does the function return, and what does name control?
The function returns a Tensor based on value. The name argument can provide a name when TensorFlow creates a new Tensor.
Follow-up: What other failures can occur? A registered converter returning an invalid result raises RuntimeError, while an incompatible tensor dtype can raise ValueError in graph mode.
Related
tf.casttf.constanttf.Tensortf.convert_to_tensor_v2