Module: tf.experimental.numpy

tf.experimental.numpy: NumPy API on TensorFlow.

This module provides a subset of NumPy API, built on top of TensorFlow operations. APIs are based on and have been tested with NumPy 1.16 version.

The set of supported APIs may be expanded over time. Also future releases may change the baseline version of NumPy API being supported. A list of some systematic differences with NumPy are listed later in the "Differences with NumPy" section.

Getting Started

Please also see TensorFlow NumPy Guide.

In the code snippets below, we will assume that tf.experimental.numpy is imported as tnp and NumPy is imported as np

print(tnp.ones([2,1]) + tnp.ones([1, 2]))


The module provides an ndarray class which wraps an immutable tf.Tensor. Additional functions are provided which accept array-like objects. Here array-like objects includes ndarrays as defined by this module, as well as tf.Tensor, in addition to types accepted by NumPy.

A subset of NumPy dtypes are supported. Type promotion follows NumPy semantics.

print(tnp.ones([1, 2], dtype=tnp.int16) + tnp.ones([2, 1], dtype=tnp.uint8))

Array Interface

The ndarray class implements the __array__ interface. This should allow these objects to be passed into contexts that expect a NumPy or array-like object (e.g. matplotlib).

np.sum(tnp.ones([1, 2]) + np.ones([2, 1]))

TF Interoperability

The TF-NumPy API calls can be interleaved with TensorFlow calls without incurring Tensor data copies. This is true even if the ndarray or tf.Tensor is placed on a non-CPU device.

In general, the expected behavior should be on par with that of code involving tf.Tensor and running stateless TensorFlow functions on them.

tnp.sum(tnp.ones([1, 2]) + tf.ones([2, 1]))

Note that the __array_priority__ is currently chosen to be lower than tf.Tensor. Hence the + operator above returns a tf.Tensor.

Additional examples of interopability include:

  • using with tf.GradientTape() scope to compute gradients through the TF-NumPy API calls.
  • using tf.distribution.Strategy scope for distributed execution
  • using tf.vectorized_map() for speeding up code using auto-vectorization

Device Support

Given that ndarray and functions wrap TensorFlow constructs, the code will have GPU and TPU support on par with TensorFlow. Device placement can be controlled by using with tf.device scopes. Note that these devices could be local or remote.

with tf.device("GPU:0"):
  x = tnp.ones([1, 2])

Graph and Eager Modes

Eager mode execution should typically match NumPy semantics of executing op-by-op. However the same code can be executed in graph mode, by putting it inside a tf.function. The function body can contain NumPy code, and the inputs can be ndarray as well.

def f(x, y):
  return tnp.sum(x + y)

f(tnp.ones([1, 2]), tf.ones([2, 1]))

Python control flow based on ndarray values will be translated by autograph into tf.cond and tf.while_loop constructs. The code can be XLA compiled for further optimizations.

However, note that graph mode execution can change behavior of certain operations since symbolic execution may not have information that is computed during runtime. Some differences are:

  • Shapes can be incomplete or unknown in graph mode. This means that ndarray.shape, ndarray.size and ndarray.ndim can return ndarray objects instead of returning integer (or tuple of integer) values.
  • __len__, __iter__ and __index__ properties of ndarray may similarly not be supported in graph mode. Code using these may need to change to explicit shape operations or control flow constructs.
  • Also note the autograph limitations.

Mutation and Variables

ndarrays currently wrap immutable tf.Tensor. Hence mutation operations like slice assigns are not supported. This may change in the future. Note however that one can directly construct a tf.Variable and use that with the TF-NumPy APIs.

tf_var = tf.Variable(2.0)

Differences with NumPy

Here is a non-exhaustive list of differences:

  • Not all dtypes are currently supported. e.g. np.float96, np.float128. np.object, np.str, np.recarray types are not supported.
  • ndarray storage is in C order only. Fortran order, views, stride_tricks are not supported.
  • Only a subset of functions and modules are supported. This set will be expanded over time. For supported functions, some arguments or argument values may not be supported. This differences are generally provide in the function comments. Full ufunc support is also not provided.
  • Buffer mutation is currently not supported. ndarrays wrap immutable tensors. This means that output buffer arguments (e..g out in ufuncs) are not supported
  • NumPy C API is not supported. NumPy's Cython and Swig integration are not supported.


random module: Public API for tf.experimental.numpy.random namespace.


class bool_: Boolean type (True or False), stored as a byte.

class complex128: Complex number type composed of two double-precision floating-point

class complex64: Complex number type composed of two single-precision floating-point

class complex_: Complex number type composed of two double-precision floating-point

class float16: Half-precision floating-point number type.

class float32: Single-precision floating-point number type, compatible with C float.

class float64: Double-precision floating-point number type, compatible with Python float

class float_: Double-precision floating-point number type, compatible with Python float

class iinfo: iinfo(type)

class inexact: Abstract base class of all numeric scalar types with a (potentially)

class int16: Signed integer type, compatible with C short.

class int32: Signed integer type, compatible with C int.

class int64: Signed integer type, compatible with Python int anc C long.

class int8: Signed integer type, compatible with C char.

class int_: Signed integer type, compatible with Python int anc C long.

class ndarray: Equivalent of numpy.ndarray backed by TensorFlow tensors.

class object_: Any Python object.

class string_: bytes(iterable_of_ints) -> bytes

class uint16: Unsigned integer type, compatible with C unsigned short.

class uint32: Unsigned integer type, compatible with C unsigned int.

class uint64: Unsigned integer type, compatible with C unsigned long.

class uint8: Unsigned integer type, compatible with C unsigned char.

class unicode_: str(object='') -> str


abs(...): TensorFlow variant of NumPy's abs.

absolute(...): TensorFlow variant of NumPy's absolute.

add(...): TensorFlow variant of NumPy's add.

all(...): TensorFlow variant of NumPy's all.

allclose(...): TensorFlow variant of NumPy's allclose.

amax(...): TensorFlow variant of NumPy's amax.

amin(...): TensorFlow variant of NumPy's amin.

angle(...): TensorFlow variant of NumPy's angle.

any(...): TensorFlow variant of NumPy's any.

append(...): TensorFlow variant of NumPy's append.

arange(...): TensorFlow variant of NumPy's arange.

arccos(...): TensorFlow variant of NumPy's arccos.

arccosh(...): TensorFlow variant of NumPy's arccosh.

arcsin(...): TensorFlow variant of NumPy's arcsin.

arcsinh(...): TensorFlow variant of NumPy's arcsinh.

arctan(...): TensorFlow variant of NumPy's arctan.

arctan2(...): TensorFlow variant of NumPy's arctan2.

arctanh(...): TensorFlow variant of NumPy's arctanh.

argmax(...): TensorFlow variant of NumPy's argmax.

argmin(...): TensorFlow variant of NumPy's argmin.

argsort(...): TensorFlow variant of NumPy's argsort.

around(...): TensorFlow variant of NumPy's around.