meridian.backend.numpy_function

Wraps a python function and uses it as a TensorFlow op.

meridian.backend.numpy_function(
    func=None, inp=None, Tout=None, stateful=True, name=None
)

Given a python function func wrap this function as an operation in a tf.function. func must take numpy arrays as its arguments and return numpy arrays as its outputs.

There are two ways to use tf.numpy_function.

As a decorator

When using tf.numpy_function as a decorator:

• you must set Tout
• you may set name
• you must not set func or inp

>>> @tf.numpy_function(Tout=tf.float32)
... def my_numpy_func(x):
...   # x will be a numpy array with the contents of the input to the
...   # tf.function
...   print(f'executing eagerly, {x=}')
...   return np.sinh(x)

The function runs eagerly:

>>> my_numpy_func(1.0).numpy()
executing eagerly, x=1.0
1.17520

The behavior doesn't change inside a tf.function:

>>> @tf.function(input_signature=[tf.TensorSpec(None, tf.float32)])
... def tf_function(input):
...   y = tf.numpy_function(my_numpy_func, [input], tf.float32)
...   return y
>>> tf_function(tf.constant(1.)).numpy()
executing eagerly, x=array(1.)
1.17520

Inplace

This form can be useful if you don't control the function's source, but it is harder to read.

Here is the same function with no decorator:

>>> def my_func(x):
...   # x will be a numpy array with the contents of the input to the
...   # tf.function
...   print(f'executing eagerly, {x=}')
...   return np.sinh(x)

To run tf.numpy_function in-place, pass the function, its inputs, and the output type in a single call to tf.numpy_function:

>>> tf.numpy_function(my_func, [tf.constant(1.0)], tf.float32)
executing eagerly, x=array(1.)
1.17520

More info

Comparison to tf.py_function: tf.py_function and tf.numpy_function are very similar, except that tf.numpy_function takes numpy arrays, and not tf.Tensors. If you want the function to contain tf.Tensors, and have any TensorFlow operations executed in the function be differentiable, please use tf.py_function.

• Calling tf.numpy_function will acquire the Python Global Interpreter Lock (GIL) that allows only one thread to run at any point in time. This will preclude efficient parallelization and distribution of the execution of the program. Therefore, you are discouraged to use tf.numpy_function outside of prototyping and experimentation.

• The body of the function (i.e. func) will not be serialized in a tf.SavedModel. Therefore, you should not use this function if you need to serialize your model and restore it in a different environment.

• The operation must run in the same address space as the Python program that calls tf.numpy_function(). If you are using distributed TensorFlow, you must run a tf.distribute.Server in the same process as the program that calls tf.numpy_function you must pin the created operation to a device in that server (e.g. using with tf.device():).

• Currently tf.numpy_function is not compatible with XLA. Calling tf.numpy_function inside tf.function(jit_compile=True) will raise an error.

• Since the function takes numpy arrays, you cannot take gradients through a numpy_function. If you require something that is differentiable, please consider using tf.py_function.