This package adds cleaner typing for numpy arrays. It has been designed for deep learning and data processing tasks which generally need a lot of numpy arrays of large dimensions.

To use the library modify all your import with :

from numpy_typing import np, ax

The modified version of numpy imported there contain sumplementary annotations in order to have smart and automatic inferred annotation.

Remark : The numpy imported there it is not slower than the classical numpy (import numpy as np). When calling a numpy function, its code is directly called without any wrapper. Numpy-typing only overload numpy's function annotations.

You can then use new annotation like:

from numpy_typing import np, ax

float32Array3d:np.float32_3d[ax.batch, ax.sample, ax.feature] = np.zeros((3, 3, 3))
v = float32Array3d[0, 0, 0] # v automatically inferred as float32

The library add new array_types annotation:

• np.float32_(n)d
• np.float64_(n)d
• np.int32_(n)d
• np.int64_(n)d
• np.int8_(n)d
• np.bool_(n)d
• np.str_(n)d

Where n is the number of dimension of the array. The value of n is for instance only supported between [1, 4].

Moreover, as the library dosen't support yet all the numpy types, you can also use the generic type np.array_(1-4)d[dtype, axis1, ...] without type.

Then you should specify for each dimension the role of the dimension:

• ax.batch: The axis select the nth batch
• ax.sample: The axis select the nth sample of the batch
• ax.feature: The axis contain features
• ax.time: The axis represent the time
• ax.label: the axis contain labels
• ax.x: The axis represent the x coordinate
• ax.y: The axis represent the y coordinate
• ax.z: The axis represent the z coordinate
• ax.rgb: the axis contain a rgb value [0] for red, [1] for green and [2] for blue
• ax.rgba: the axis contain a rgba value [0] for red, [1] for green, [2] for blue and [3] for alpha

Here is an example of useful smart annotation:

from numpy_typing import np, ax

a:np.float32_1d[ax.time] = np.zeros((64))
b:np.float32_1d[ax.time] = np.zeros((64))
c = np.concatenate([a, b])
# automatically infer the type of c as np.float32_1d[ax.time]

If you want the library to support more types, more numpy functions or if you have any suggestion, feel free to open an issue on our github.