kde_gpu

Kernel density estimator using Nadaraya-Watson with GPU (CUDA)

Author

Chen Chen

Running Environment Setup

You have to have cupy installed to be able to use GPU!! See https://github.com/cupy/cupy

Similar to scipy.kde_gaussian and statsmodels.nonparametric.kernel_density.KDEMultivariateConditional, this project implements Nadaraya-Waston kernel density and kernel conditional probability estimator using cuda through cupy. However, it is much faster than CPU version and it maximise the use of GPU memory.

File tree structure:
|--- kde_gpu\
***|--- version.py \
***|--- nadaraya_watson.py \
***|--- conditional_probability.py \
*|--- setup.py \
*|--- example.py \
*|--- README.md \

1. Install cupy (i.e. numpy equivalent using CUDA GPU acceleration) according to https://docs-cupy.chainer.org/en/stable/install.html#using-pip

2. Install kde_gpu with following command: (Please use pip3 if the default python in your computer is python2)

   $ pip install kde_gpu
   

This command will run setup.py where we specify the dependencies required to run kde_gpu. The dependencies we require are:

       "scipy>=1.0.0",
       "pandas>=0.20.2",

Please note that if the version number of installed package in your machine is lower than the stated version number, pip will uninstall your out-of-date package and install the one with version number greater than or equal to the stated one in setup.py.

Example

"""
@author: chen.chen.adl@gmail.com
"""

#import kernel_smoothing
from scipy import stats
import pandas as pd
import cupy as cp
import numpy as np
import time

#true distribution is an exponential
rv = stats.expon(0,1)
x = rv.rvs(size=10000)

density_real = rv.pdf(x)

#using scipy package
t1=time.time()
kde_scipy=stats.gaussian_kde(x.T,bw_method='silverman')
kde_scipy=kde_scipy(x.T)
print(time.time()-t1)

#use kde_gpu package
t1=time.time()
kde_cupy=kde(cp.asarray(x.T),bw_method='silverman')
print(time.time()-t1)

df = pd.DataFrame({'x1':x,'kde_scipy':kde_scipy,
                   'kde_cupy':cp.asnumpy(kde_cupy).squeeze(),'real density':density_real})

df['scipy_mean_absolute_error']=np.abs(df['kde_scipy']-df['real density'])
df['cupy_mean_absolute_error']=np.abs(df['kde_cupy']-df['real density'])
print(df.mean())


#true x is truncated normal and true y conditional on x is a uniform distribution, which is independent of x.
rv = stats.truncnorm(-3,2,30,10)
nsample=10000
x = cp.asarray(rv.rvs(nsample))
ycondx = cp.asarray(cp.random.rand(nsample))
y = 10*(ycondx-0.5)+x

cdf_conditional_real = ycondx
df = pd.DataFrame({'y':cp.asnumpy(y),'x':cp.asnumpy(x),'real density':cp.asnumpy(cdf_conditional_real)})

df['nadaraya watson']= kernel_smoothing_ecdf(y,x)
df['nw_error']=np.abs(df['nadaraya watson']-df['real density'])
df.mean()