1. Quick Start¶
1.1. Install Khiva¶
First of all, the Khiva C++ library should be installed:
Then, install the compiled Khiva package that is hosted on the Python Package Index (PyPI) with pip:
pip install khiva
1.2. Dive in¶
Dive quickly into Khiva with the following example: First, set the backend and device you want to use. There is a backend and a device set by default:
from khiva.library import *
set_backend(KHIVABackend.KHIVA_BACKEND_OPENCL)
set_device(0)
Then, you can create an array in the device:
from khiva.array import *
a = Array.from_list([1, 2, 3, 4, 5, 6, 7, 8], dtype.s32)
a.display()
The previous lines print the dimensions and the content of the created array:
| [8 1 1 1] |
|---|
| 1.0000 |
| 2.0000 |
| 3.0000 |
| 4.0000 |
| 5.0000 |
| 6.0000 |
| 7.0000 |
| 8.0000 |
Once the array is created in device memory, we can concatenate operations with this array in an asynchronous way and receive the data only in the host when to_list(), to_numpy() or to_pandas() (the latter only supports bi-dimensional time series) functions are called.
a = a.to_pandas()
print(a)
The result is the next one:
| 0 | |
|---|---|
| 1 | 1.0 |
| 2 | 2.00 |
| 3 | 3.00 |
| 4 | 4.00 |
| 5 | 5.00 |
| 6 | 6.00 |
| 7 | 7.00 |
| 8 | 8.00 |
Now, let’s dive into the asynchronous usage of the library. Khiva library provides us several time series analysis functionalities which include features extraction, time-series re-dimension, distance calculations, motifs and discords detection, tools for similarity study, statistical parameters extraction or time series normalization.
All these functionalities can be concatenated to improve the performance, so you can get the data just in the moment that you do not use the functions of this library:
from khiva.matrix import *
stomp_result = stomp(Array.from_numpy(np.array([11, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 11]), dtype.s32),
Array.from_numpy(np.array([9, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 9]), dtype.s32),
3)
find_best_n_discords_result = find_best_n_discords(stomp_result[0],
stomp_result[1], 2)
a = find_best_n_discords_result[2].to_numpy()
print(a)
The previous produces the following output:
| [1.73190141 1.73185158] [8 8] [0 9] |
The first numpy array represents the minimum distances between the subsequences of length 3 between the two time-series. The second numpy array represents the location of those subsequences in the first time-series and the third one represents the indices in the second time-series.
We want to highlight the possibility of using the library for computing the functions in different backends and with different devices, knowing that the operations should be executed in the same device where the array was created.
#Adding operations in the different backends and devices.
from khiva.features import *
set_backend(KHIVABackend.KHIVA_BACKEND_OPENCL)
set_device(0)
a = Array.from_list([1, 2, 3, 4, 5, 6, 7, 8], dtype.s32)
b = mean(a)
set_device(1)
c = Array.from_list([1, 2, 3, 4, 5, 6, 7, 8], dtype.s32)
d = mean(c)
set_backend(KHIVABackend.KHIVA_BACKEND_CPU)
set_device(0)
e = Array([1, 2, 3, 4, 5, 6, 7, 8])
f = mean(e)
#Retrieving the results of the previous operations
set_backend(KHIVABackend.KHIVA_BACKEND_OPENCL)
set_device(0)
print(b.to_numpy())
set_device(1)
print(d.to_numpy())
set_backend(KHIVABackend.KHIVA_BACKEND_CPU)
set_device(0)
print(f.to_numpy())
The output is the next one:
| 4.5 |
| 4.5 |
| 4.5 |
Note that the data type used by default is floating point of 32 bits in order to avoid problems with the different devices, but it can be changed deliberately.
The available data types are the next ones:
| Data type | Explanation |
|---|---|
| f32 | 32 bits Float |
| c32 | 32 bits Complex |
| f64 | 64 bits Double |
| c64 | 64 bits Complex |
| b8 | 8 bits Boolean |
| s32 | 32 bits Int |
| 32u | 32 bits Unsigned Int |
| u8 | 8 bits Unsigned Int |
| s64 | 64 bits Int |
| u64 | 64 bits Unsigned Int |
| s16 | 16 bits Int |
| u16 | 16 bits Unsigned Int |
There are functions that do not support 32 bits floating point data type, so it is necessary to indicate the data type. The following is an example function requiring a 32-bit signed integer array:
cwt_coefficients_result = cwt_coefficients(Array.from_list([[0.1, 0.2, 0.3], [0.1, 0.2, 0.3]], dtype.s32),
Array.from_list([1, 2, 3], dtype.s32), 2, 2).to_numpy()
print(cwt_coefficients_result)
The output is:
| [0.26517162 0.26517162] |
1.3. Limitations¶
This open-source library provides a very good performance, but it has got memory limitations. For cases where you need to apply a time series analysis over a huge amount of data and in short-term fashion, please, contact us).
1.4. Let’s Rock!¶
Now, you have the basic concepts to start using the library. Please, follow the documentation of each function to know how to use them. Each function has its corresponding tests so you can check how to use each of them.
Furthermore, we provide use cases and examples that you can use to learn where and how to apply the library.