Is vectorizing this triple for loop in Python / Numpy possible? - python

I am trying to speed up my code which currently takes a little over an hour to run in Python / Numpy. The majority of computation time occurs in the function pasted below.
I'm trying to vectorize Z, but I'm finding it rather difficult for a triple for loop. Could I possible implement the numpy.diff function somewhere? Take a look:
def MyFESolver(KK,D,r,Z):
global tdim
global xdim
global q1
global q2
for k in range(1,tdim):
for i in range(1,xdim-1):
for j in range (1,xdim-1):
Z[k,i,j]=Z[k-1,i,j]+r*q1*Z[k-1,i,j]*(KK-Z[k-1,i,j])+D*q2*(Z[k-1,i-1,j]-4*Z[k-1,i,j]+Z[k-1,i+1,j]+Z[k-1,i,j-1]+Z[k-1,i,j+1])
return Z
tdim = 75 xdim = 25

I agree, it's tricky because the BCs on all four sides, ruin the simple structure of the Stiffness matrix. You can get rid of the space loops as such:
from pylab import *
from scipy.sparse.lil import lil_matrix
tdim = 3; xdim = 4; r = 1.0; q1, q2 = .05, .05; KK= 1.0; D = .5 #random values
Z = ones((tdim, xdim, xdim))
#Iterate in time
for k in range(1,tdim):
Z_prev = Z[k-1,:,:] #may need to flatten
Z_up = Z_prev[1:-1,2:]
Z_down = Z_prev[1:-1,:-2]
Z_left = Z_prev[:-2,1:-1]
Z_right = Z_prev[2:,1:-1]
centre_term = (q1*r*(Z_prev[1:-1,1:-1] + KK) - 4*D*q2)* Z_prev[1:-1,1:-1]
Z[k,1:-1,1:-1]= Z_prev[1:-1,1:-1]+ centre_term + q2*(Z_up+Z_left+Z_right+Z_down)
But I don't think you can get rid of the time loop...
I think the expression:
Z_up = Z_prev[1:-1,2:]
makes a copy in numpy, whereas what you want is a view - if you can figure out how to do this - it should be even faster (how much?)
Finally, I agree with the rest of the answerers - from experience, this kind of loops are better done in C and then wrapped into numpy. But the above should be faster than the original...

This looks like an ideal case for Cython. I'd suggest writing that function in Cython, it'll probably be hundreds of times faster.

Related

Parallelizing for loops using Dask (or other efficient way)

I have a function who takes xarray data set (similar to pandas multi-index) and uses 4 for loops embedded in each other to compute a new data array variable.
I wonder if there is a way I can use Dask to make this process faster, I'm quite new to this so I'm not sure.
The function looks like this:
def A_calc(data, thresh):
A = np.zeros((len(data.time), len(data.lat), len(data.lon)))
foo = xr.DataArray(A, coords=[data.time, data.lat, data.lon],
dims=['time','lat', 'lon'])
for t in tqdm(range(len(data.time))):
for i in range(len(data.lat)):
for j in range(2,len(data.lon)):
for k in range(len(data.lev)):
if np.isnan(
data[dict(time=[t], lat=[i], lon=[j], lev=[k])].sigma_0.values):
foo[dict(time=[t], lat=[i], lon=[j])] = np.nan
break
elif abs(
data[dict(time=[t], lat=[i], lon=[j], lev=[k])].sigma_0.values
- data[dict(time=[t], lat=[i], lon=[j], lev=[1])].sigma_0.values) >= thresh:
foo[dict(time=[t], lat=[i], lon=[j])] = data.lev[k].values
break
return foo
Any suggestions?
As is said in the comments, Python for loops are slow. Typically the first step to accelerating code like to this is to either ...
Find some clever way to write all of this as a vectorized numpy expression, without Python for loops
Use Numba

Vectorize python code for improved performance

I am writing a scientific code in python to calculate the energy of a system.
Here is my function : cte1, cte2, cte3, cte4 are constants previously computed; pii is np.pi (calculated beforehand, since it slows the loop otherwise). I calculate the 3 components of the total energy, then sum them up.
def calc_energy(diam):
Energy1 = cte2*((pii*diam**2/4)*t)
Energy2 = cte4*(pii*diam)*t
d=diam/t
u=np.sqrt((d)**2/(1+d**2))
cc= u**2
E = sp.special.ellipe(cc)
K = sp.special.ellipk(cc)
Id=cte3*d*(d**2+(1-d**2)*E/u-K/u)
Energy3 = cte*t**3*Id
total_energy = Energy1+Energy2+Energy3
return (total_energy,Energy1)
My first idea was to simply loop over all values of the diameter :
start_diam, stop_diam, step_diam = 1e-10, 500e-6, 1e-9 #Diametre
diametres = np.arange(start_diam,stop_diam,step_diam)
for d in diametres:
res1,res2 = calc_energy(d)
totalEnergy.append(res1)
Energy1.append(res2)
In an attempt to speed up calculations, I decided to use numpy to vectorize, as shown below :
diams = diametres.reshape(-1,1) #If not reshaped, calculations won't run
r1 = np.apply_along_axis(calc_energy,1,diams)
However, the "vectorized" solution does not properly work. When timing I get 5 seconds for the first solution and 18 seconds for the second one.
I guess I'm doing something the wrong way but can't figure out what.
With your current approach, you're applying a Python function to each element of your array, which carries additional overhead. Instead, you can pass the whole array to your function and get an array of answers back. Your existing function appears to work fine without any modification.
import numpy as np
from scipy import special
cte = 2
cte1 = 2
cte2 = 2
cte3 = 2
cte4 = 2
pii = np.pi
t = 2
def calc_energy(diam):
Energy1 = cte2*((pii*diam**2/4)*t)
Energy2 = cte4*(pii*diam)*t
d=diam/t
u=np.sqrt((d)**2/(1+d**2))
cc= u**2
E = special.ellipe(cc)
K = special.ellipk(cc)
Id=cte3*d*(d**2+(1-d**2)*E/u-K/u)
Energy3 = cte*t**3*Id
total_energy = Energy1+Energy2+Energy3
return (total_energy,Energy1)
start_diam, stop_diam, step_diam = 1e-10, 500e-6, 1e-9 #Diametre
diametres = np.arange(start_diam,stop_diam,step_diam)
a = calc_energy(diametres) # Pass the whole array

Python curve fit with change point

As I'm really struggleing to get from R-code, to Python code, I would like to ask some help. The code I want to use has been provided to my from withing the mathematics forum of stackexchange.
https://math.stackexchange.com/questions/2205573/curve-fitting-on-dataset
I do understand what is going on. But I'm really having a hard time trying to solve the R-code, as I have never seen anything of it. I have written the function to return the sum of squares. But I'm stuck at how I could use a function similar to the optim function. And also I don't really like the guesswork at the initial values. I would like it better to run and re-run a type of optim function untill I get the wanted result, because my needs for a nearly perfect curve fit are really high.
def model (par,x):
n = len(x)
res = []
for i in range(1,n):
A0 = par[3] + (par[4]-par[1])*par[6] + (par[5]-par[2])*par[6]**2
if(x[i] == par[6]):
res[i] = A0 + par[1]*x[i] + par[2]*x[i]**2
else:
res[i] = par[3] + par[4]*x[i] + par[5]*x[i]**2
return res
This is my model function...
def sum_squares (par, x, y):
ss = sum((y-model(par,x))^2)
return ss
And this is the sum of squares
But I have no idea on how to convert this:
#I found these initial values with a few minutes of guess and check.
par0 <- c(7,-1,-395,70,-2.3,10)
sol <- optim(par= par0, fn=sqerror, x=x, y=y)$par
To Python code...
I wrote an open source Python package (BSD license) that has a genetic algorithm (Differential Evolution) front end to the scipy Levenberg-Marquardt solver, it functions similarly to what you describe in your question. The github URL is:
https://github.com/zunzun/pyeq3
It comes with a "user-defined function" example that's fairly easy to use:
https://github.com/zunzun/pyeq3/blob/master/Examples/Simple/FitUserDefinedFunction_2D.py
along with command-line, GUI, cluster, parallel, and web-based examples. You can install the package with "pip3 install pyeq3" to see if it might suit your needs.
Seems like I have been able to fix the problem.
def model (par,x):
n = len(x)
res = np.array([])
for i in range(0,n):
A0 = par[2] + (par[3]-par[0])*par[5] + (par[4]-par[1])*par[5]**2
if(x[i] <= par[5]):
res = np.append(res, A0 + par[0]*x[i] + par[1]*x[i]**2)
else:
res = np.append(res,par[2] + par[3]*x[i] + par[4]*x[i]**2)
return res
def sum_squares (par, x, y):
ss = sum((y-model(par,x))**2)
print('Sum of squares = {0}'.format(ss))
return ss
And then I used the functions as follow:
parameter = sy.array([0.0,-8.0,0.0018,0.0018,0,200])
res = least_squares(sum_squares, parameter, bounds=(-360,360), args=(x1,y1),verbose = 1)
The only problem is that it doesn't produce the results I'm looking for... And that is mainly because my x values are [0,360] and the Y values only vary by about 0.2, so it's a hard nut to crack for this function, and it produces this (poor) result:
Result
I think that the range of x values [0, 360] and y values (which you say is ~0.2) is probably not the problem. Getting good initial values for the parameters is probably much more important.
In Python with numpy / scipy, you would definitely want to not loop over values of x but do something more like
def model(par,x):
res = par[2] + par[3]*x + par[4]*x**2
A0 = par[2] + (par[3]-par[0])*par[5] + (par[4]-par[1])*par[5]**2
res[np.where(x <= par[5])] = A0 + par[0]*x + par[1]*x**2
return res
It's not clear to me that that form is really what you want: why should A0 (a value independent of x added to a portion of the model) be so complicated and interdependent on the other parameters?
More importantly, your sum_of_squares() function is actually not what least_squares() wants: you should return the residual array, you should not do the sum of squares yourself. So, that should be
def sum_of_squares(par, x, y):
return (y - model(par, x))
But most importantly, there is a conceptual problem that is probably going to plague this model: Your par[5] is meant to represent a breakpoint where the model changes form. This is going to be very hard for these optimization routines to find. These routines generally make a very small change to each parameter value to estimate to derivative of the residual array with respect to that variable in order to figure out how to change that variable. With a parameter that is essentially used as an integer, the small change in the initial value will have no effect at all, and the algorithm will not be able to determine the value for this parameter. With some of the scipy.optimize algorithms (notably, leastsq) you can specify a scale for the relative change to make. With leastsq that is called epsfcn. You may need to set this as high as 0.3 or 1.0 for fitting the breakpoint to work. Unfortunately, this cannot be set per variable, only per fit. You might need to experiment with this and other options to least_squares or leastsq.

How do I vectorize the following loop in Numpy?

"""Some simulations to predict the future portfolio value based on past distribution. x is
a numpy array that contains past returns.The interpolated_returns are the returns
generated from the cdf of the past returns to simulate future returns. The portfolio
starts with a value of 100. portfolio_value is filled up progressively as
the program goes through every loop. The value is multiplied by the returns in that
period and a dollar is removed."""
portfolio_final = []
for i in range(10000):
portfolio_value = [100]
rand_values = np.random.rand(600)
interpolated_returns = np.interp(rand_values,cdf_values,x)
interpolated_returns = np.add(interpolated_returns,1)
for j in range(1,len(interpolated_returns)+1):
portfolio_value.append(interpolated_returns[j-1]*portfolio_value[j-1])
portfolio_value[j] = portfolio_value[j]-1
portfolio_final.append(portfolio_value[-1])
print (np.mean(portfolio_final))
I couldn't find a way to write this code using numpy. I was having a look at iterations using nditer but I was unable to move ahead with that.
I guess the easiest way to figure out how you can vectorize your stuff would be to look at the equations that govern your evolution and see how your portfolio actually iterates, finding patterns that could be vectorized instead of trying to vectorize the code you already have. You would have noticed that the cumprod actually appears quite often in your iterations.
Nevertheless you can find the semi-vectorized code below. I included your code as well such that you can compare the results. I also included a simple loop version of your code which is much easier to read and translatable into mathematical equations. So if you share this code with somebody else I would definitely use the simple loop option. If you want some fancy-pants vectorizing you can use the vector version. In case you need to keep track of your single steps you can also add an array to the simple loop option and append the pv at every step.
Hope that helps.
Edit: I have not tested anything for speed. That's something you can easily do yourself with timeit.
import numpy as np
from scipy.special import erf
# Prepare simple return model - Normal distributed with mu &sigma = 0.01
x = np.linspace(-10,10,100)
cdf_values = 0.5*(1+erf((x-0.01)/(0.01*np.sqrt(2))))
# Prepare setup such that every code snippet uses the same number of steps
# and the same random numbers
nSteps = 600
nIterations = 1
rnd = np.random.rand(nSteps)
# Your code - Gives the (supposedly) correct results
portfolio_final = []
for i in range(nIterations):
portfolio_value = [100]
rand_values = rnd
interpolated_returns = np.interp(rand_values,cdf_values,x)
interpolated_returns = np.add(interpolated_returns,1)
for j in range(1,len(interpolated_returns)+1):
portfolio_value.append(interpolated_returns[j-1]*portfolio_value[j-1])
portfolio_value[j] = portfolio_value[j]-1
portfolio_final.append(portfolio_value[-1])
print (np.mean(portfolio_final))
# Using vectors
portfolio_final = []
for i in range(nIterations):
portfolio_values = np.ones(nSteps)*100.0
rcp = np.cumprod(np.interp(rnd,cdf_values,x) + 1)
portfolio_values = rcp * (portfolio_values - np.cumsum(1.0/rcp))
portfolio_final.append(portfolio_values[-1])
print (np.mean(portfolio_final))
# Simple loop
portfolio_final = []
for i in range(nIterations):
pv = 100
rets = np.interp(rnd,cdf_values,x) + 1
for i in range(nSteps):
pv = pv * rets[i] - 1
portfolio_final.append(pv)
print (np.mean(portfolio_final))
Forget about np.nditer. It does not improve the speed of iterations. Only use if you intend to go one and use the C version (via cython).
I'm puzzled about that inner loop. What is it supposed to be doing special? Why the loop?
In tests with simulated values these 2 blocks of code produce the same thing:
interpolated_returns = np.add(interpolated_returns,1)
for j in range(1,len(interpolated_returns)+1):
portfolio_value.append(interpolated_returns[j-1]*portfolio[j-1])
portfolio_value[j] = portfolio_value[j]-1
interpolated_returns = (interpolated_returns+1)*portfolio - 1
portfolio_value = portfolio_value + interpolated_returns.tolist()
I assuming that interpolated_returns and portfolio are 1d arrays of the same length.

Numba function slower than C++ and loop re-order further slows down x10

The following code simulates extracting binary words from different locations within a set of images.
The Numba wrapped function, wordcalc in the code below, has 2 problems:
It is 3 times slower compared to a similar implementation in C++.
Most strangely, if you switch the order of the "ibase" and "ibit" for-loops, speed drops by a factor of 10 (!). This does not happen in the C++ implementation which remains unaffected.
I'm using Numba 0.18.2 from WinPython 2.7
What could be causing this?
imDim = 80
numInsts = 10**4
numInstsSub = 10**4/4
bitsNum = 13;
Xs = np.random.rand(numInsts, imDim**2)
iInstInds = np.array(range(numInsts)[::4])
baseInds = np.arange(imDim**2 - imDim*20 + 1)
ofst1 = np.random.randint(0, imDim*20, bitsNum)
ofst2 = np.random.randint(0, imDim*20, bitsNum)
#nb.jit(nopython=True)
def wordcalc(Xs, iInstInds, baseInds, ofst, bitsNum, newXz):
count = 0
for i in iInstInds:
Xi = Xs[i]
for ibit in range(bitsNum):
for ibase in range(baseInds.shape[0]):
u = Xi[baseInds[ibase] + ofst[0, ibit]] > Xi[baseInds[ibase] + ofst[1, ibit]]
newXz[count, ibase] = newXz[count, ibase] | np.uint16(u * (2**ibit))
count += 1
return newXz
ret = wordcalc(Xs, iInstInds, baseInds, np.array([ofst1, ofst2]), bitsNum, np.zeros((iInstInds.size, baseInds.size), dtype=np.uint16))
I get 4x speed-up by changing from np.uint16(u * (2**ibit)) to np.uint16(u << ibit); i.e. replace the power of 2 with a bitshift, which should be equivalent (for integers).
It seems reasonably likely that your C++ compiler might be making this substitution itself.
Swapping the order of the two loops makes a small difference for me for both your original version (5%) and my optimized version (15%), so I can't think I can make a useful comment on that.
If you really wanted to compare the Numba and C++ you can look at the compiled Numba function by doing os.environ['NUMBA_DUMP_ASSEMBLY']='1' before you import Numba. (That's clearly quite involved though).
For reference, I'm using Numba 0.19.1.

Categories

Resources