ZMCintegral (Numba backened) is an easy to use python package which uses Monte Carlo Evaluation Method to do numerical integrations on Multi-GPU devices. It supports integrations with up to 16 multi-variables, and it is capable of even more than 16 variables if time is not of the priori concern.
To understand how ZMCintegral works, please refer to
https://arxiv.org/pdf/1902.07916v2.pdf
This new version supports parameter grid search, for this new functionality please refer to
https://arxiv.org/pdf/1910.01965.pdf
ZMCintegral usually takes a few minutes to finish the task.
- Full flexibility of user defined functions
- Multi-dimension integration
- Multi-GPU supports
- Stratified sampling
- Heuristic tree search
- Parameter grid search
- very many functions
To run ZMCintegral (Numba-Ray version), the following packages needs to be pre-installed:
- Numba
- Ray
- cudatoolkit
# create a new environment
$: conda create -n zmcintegral
# install relavant package sequentially
$: pip install ZMCintegral
$: conda install python=3.6
$: conda install numba=0.45.1
$: conda install cudatoolkit (**notification)
$: pip install -U ray[debug]==0.7.1
** notification
make sure it is compatable with your driver verison, other wise you will get an error:
numba.cuda.cudadrv.driver.LinkerError: [218] Call to cuLinkAddData results in UNKNOWN_CUDA_ERROR
ptxas application ptx input, line 9; fatal : Unsupported .version 6.4; current version is '6.2'
ptxas fatal : Ptx assembly aborted due to errors
First of all, prepare machines with Nvidia GPU devices. choose one of them as a head node:
# for head node
$: ray start --head --redis-port=XXXX --num-cpus=10 --num-gpus=4
#for other nodes, here the redis-address is the ip of head node.
$: ray start --redis-address=XXX.XXX.XX.XX:XXXX --num-cpus=5 --num-gpus=2
Remeber to use
# for head node
$: ray stop
#for other nodes
$: ray stop
after evaluation.
# this is for installation by pip install zmcintegral
from ZMCintegral.ZMC.ZMCintegral_normal import MCintegral_normal
import math
import numpy as np
import time
start = time.time()
# user defined function
fun = """
import math
# define a device function that should be used by cuda kernel
@cuda.jit(device=True)
def fun(x):
return math.sin(x[0]+x[1]+x[2]+x[3]+x[4]+x[5]+x[6])
"""
# define arguments that MCintegral_normal requires
depth = 1
sigma_multiplier = 5
num_trials = 5
num_chunks_in_one_dimension = 12
# call MCintegral_normal
MC = MCintegral_normal(my_func = fun,
domain = [[0,10],[0,10],[0,10],[0,10],[0,10],[0,10]],
head_node_address = "XXX.XXX.XX.XX:XXXX",
depth = depth,
sigma_multiplier = sigma_multiplier,
num_trials = num_trials,
num_chunks_in_one_dimension = num_chunks_in_one_dimension)
# obtaining the result
result = MC.evaluate()
# print the formatted result
print('result = %s std = %s' % (result[0], result[1]))
print('evaluation time {}'.format(time.time()-start))
total number of GPUs: 1
140 hypercube(s) need(s) to be recalculated, to save time, try increasing sigma_multiplier.
result = -49.47563512703137 std = 1.9873890591413763
evaluation time 37.8058066368103
# this is for installation by pip install zmcintegral
from ZMCintegral.ZMC.ZMCintegral_functional import MCintegral_functional
import math
import numpy as np
import time
start = time.time()
# user defined function
fun = """
import math
# define a device function that should be used by cuda kernel
@cuda.jit(device=True)
def fun(x,para):
return math.sin(x[0]+x[1]+x[2]+x[3]+para[0]+para[1])
"""
# para contains two parameters
para = [[1,2,3,4,5],[1.1,2.2,3.1]]
# sample points is taken to 10**6
sample_points = 10**6
# the parameter grid has totally 5*3=15 points
# we choose 3 batches as an example
batch_size = 5
# call MCintegral_functional
MC = MCintegral_functional(my_func = fun,
domain = [[0,1],[0,1],[0,1],[0,1]],
parameters = para,
head_node_address = "XXX.XXX.XX.XX:XXXX",
num_points = sample_points,
batch_size = batch_size)
# obtaining the result
result = MC.evaluate()
result = print(np.reshape(result,[3,5]))
print('evaluation time {}'.format(time.time()-start))
<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
Detected total number of GPUs: 1
Total parameter grid size: 15, Each GPU cycle will cover 5 grid values, Total GPU cycles: 3
Evaluating, please wait...
[[-0.64027657 -0.81054914 -0.23614549 0.55573641 0.83644859]
[-0.8100885 -0.23610632 0.55550291 0.83625073 0.34960498]
[-0.23645369 0.55572789 0.83633015 0.34781143 -0.45968675]]
evaluation time 4.728282690048218
# this is for installation by pip install zmcintegral
from ZMCintegral.ZMC.ZMCintegral_multifunctions import MCintegral_multi_function
import math
import numpy as np
import time
start = time.time()
# user defined function and domains
funs = ["math.sin(x[0]+x[1]+x[2]+x[3])*math.tan(x[0])",\
"math.cos(x[0]+x[1]+x[2]+x[3])"]
domains = [[[0,1],[0,10],[0,1],[0,1]],\
[[0,1],[0,10],[0,1],[0,1]]]
# sample points is taken to 10**6
sample_points = 10**6
# call MCintegral_functional
MC = MCintegral_multi_function(my_funcs = funs,
domains = domains,
head_node_address = "XXX.XXX.XXX.XXX:XXXX",
num_points = sample_points)
# obtaining the result
result = MC.evaluate()
print('result {}'.format(np.array(result).flatten()))
print('evaluation time {:.4f} s'.format(time.time()-start))
Detected total number of GPUs: 1
Evaluating, please wait...
result [-0.425192 -1.64925066]
evaluation time 0.9935 s
The following four parameters can be tuned to fit special cases.
| parameter | usage | example | default |
|---|---|---|---|
| num_trials | Evaluate the integration for num_trials times. Better kept within 10. | 10 | 5 |
| depth | For importance sampling. A domain is magnified for depth times. Better kept within 3. | 3 | 2 |
| num_chunks_in_one_dimension | The number of chunks users want to set along one dimension | 10 | 4 |
| sigma_multiplier | Only domains that have very large standardand deviations (hence, very unstable) should be magnified and re-evaluated. Domains which are beyond sigma_multiplication * σ should be recalculated. | 3 | 4 |
The user defined function must be organized in string format as shown in the following example. And the function name in the string mutst be fun, something like:
# user defined function
fun = """
import math
# define a device function that should be used by cuda kernel
@cuda.jit(device=True)
def fun(x): # here the function name must be set as `fun`
return xxx
"""
One should read the documentation for the Numba package's CUDA capabilities when trying to use this package. ZMCintegral is only compatible with device functions as Numba does not support dynamic parallelism. This is important when designing the integrated function.
Issues with CUDA should first be resolved by looking at the CUDA documentation.
For further questions and technical issues, please contact us at
[email protected] (Hong-Zhong Wu 伍宏忠)
[email protected] (Jun-Jie Zhang 张俊杰)
[email protected] (Xiao-Yan Cao 曹晓岩)
The package is coded by Jun-Jie Zhang and checked by Hong-Zhong Wu of University of Science and Technology of China.
This package is free you can redistribute it and/or modify it under the terms of the Apache License Version 2.0, January 2004 (http://www.apache.org/licenses/).
File Structure
ZMCintegral
│ README.md
│ LICENSE
│ setup.py
│
└───ZMC
│ │ ZMCintegral_functional.py
│ │ ZMCintegral_normal.py
│ │ ZMCintegral_multifunctions.py
│ │ __init__.py
│
└───pics
│ multi-function.PNG
│ sin6d.PNG
│ sin6d_theoretical.PNG
│ parameter integration.PNG
│ parameter integration theoretical.PNG