0

I am slightly new to python and I am trying to convert some code.This is an approximation method. Which isn't important. In my oddev function I get returned

        c2[1:modes+1] = v* 1j
ValueError: could not broadcast input array from shape (25) into shape (25,1)

When I do this Matlab I believe it automatically casts it, and will store the complex array. The function is a getting the coefficient from a partial sine transform to do this. At first I tried storing the random matrix which just an array using np.matlib method and this had the same shape but I believe I will lose the real values of the filter when I cast it. How do I store this?

import math
import numpy as np
def quickcontmin(datain):

n = np.shape(datain)[0]
m = math.floor(n / 2)
modes = math.floor(m / 2)
addl = 20
nn = 20 * n
chi = 10 ** -13

def evenhp(xv):
    "Even high pass"
    n1 = np.shape(xv)[0]
    vx = np.array(xv[:-1])
    vx = vx[::-1]
    c1 = np.append(xv,vx)
    c1 = np.fft.fft(c1)       
    c1[0:modes-1] = 0.0
    c1[-1 - modes + 2:-1] = 0.0
    evenl = np.real(np.fft.ifft(c1))
    even = evenl[0:n1-1]
    return even

def evenhpt(xv):
    " Transpose of EvenHP"
    n1 = np.shape(xv)[0]
    xy = np.zeros((n1- 2, 1))
    c1 = np.append(xv,xy)
    c1 = np.fft.fft(c1)
    c1[0:modes-1] = 0.0
    c1[-1 - modes + 1:-1] = 0.0
    evenl = np.real(np.fft.ifft(c1))
    even = evenl[0:n1-1]
    even[1:-2] = even[1:-2] + evenl[-1:-1:n1+1]
    return even``

def evenlp(xv):
    " Low pass cosine filter"
    n1 = np.shape(xv)[0]
    vx = np.array(xv[:-1])
    vx = vx[::-1]
    c1 = np.append(xv,vx)
    c1 = np.fft.fft(c1)
    c1[modes + 1:-1 - modes + 1] = 0.0
    evenl = np.real(np.fft.ifft(c1))
    even = evenl[0:n1-1]
    return even

def oddev(xv):
    "Evaluate the sine modes on the grid"
    c2 = np.zeros((2 *n - 2, 1))*1j
    v = np.array(xv[:])
    v1 = v[:-1]
    v1 = v[::-1]
    c2[1:modes+1] = v* 1j
    c2[-1 - modes + 1:-1] = -v1* 1j
    evall = np.fft.ifft(c2) * math.sqrt(2 * n - 2)
    eva = evall[0:n-1]
    return eva

def oddevt(xv):
    " Transpose the sine modes on the function OddEv"
    c1 = np.array(xv[1:-2])
    c1 = np.insert(c1,0.0,0)
    c1 = np.append(c1,0.0)
    c1 = np.append(c1,xv[-2:-1:2])
    c1a = np.divide(np.fft.fft(c1),math.sqrt(2 * n - 2))
    fcoef = np.imag(c1a[1:modes])
    return fcoef

def eextnd(xv):
    "Obtain cosine coefficients and evalue on the refined grid"  
    vx = np.array(xv[:-1])
    vx = vx[::-1]
    c1  = np.append(xv,vx)
    c1 = np.fft.fft(c1)
    cL = np.zeros((2*nn-2,1))
    cL[0:modes-1] = c1[0:modes-1]
    cL[-1 - modes + 1:-1] = c1[-1 - modes + 1:-1]
    evenexL = np.multiply(np.fft.ifft(cL) , (nn - 1) / (n - 1))
    evenex = evenexL[0:nn-1]
    return evenex

def oextnd(xv):
    "Evaluate sine coefficients on the refined grid"
    c2 = np.zeros((2 * nn - 2, 1))
    c2[0] = 0.0
    c2[1:modes + 1] = np.multiply(xv[0:-1],1j)
    c2[-1 - modes + 1:-1] = np.multiply(-xv[-1:-1:1],1j)
    evall = np.real(np.multiply(np.fft.ifft(c2), math.sqrt(2 * n - 2) * (2 *nn - 2) / (2 * n - 2)))
    oox = evall[0:nn-1]
    return oox

dc = evenlp(datain)
#L in paper, number of vectors used to sample the columnspace
lll = round(4 * math.log(m )/ math.log(2)) + addl
lll = int(lll)
#The following should be straightforward from the psuedo-code
w=2 * np.random.rand(modes , lll) - 1 
p=np.matlib.zeros(shape=(n,lll))
for j in range(lll):
    p[:,j] = evenhp(oddev(w[:,j]))
q,r = np.linalg.qr(p , mode='reduced')
z = np.zeros(shape=(modes,lll))
for j in range(lll):
    z[:,j]= oddevt(evenhpt(q[:,j]))
un,s,v = np.linalg.svd(z,full_matrices='False')
ds=np.diag(s)
aa=np.extract(np.diag(s)>(chi))
aa[-1] = aa
aa = int(aa)
s = 0 * s
for j in range(aa):
    s[j,j] = 1.0 / ds(j)
#find the sine coefficents
b=un*s* v.T* q.T* evenhp(datain)
#Constructing the continuation
exs=oddev(b)
pexs = evenlp(exs)

dataCont=exs-pexs+dc
dataCont[n+1:2*n-2]=-exs[-2:-1:1]-pexs[-2:-1:1]+dc[-2:-1:1]
#Evaluate the continuation on the refined grid
dataRefined=eextnd(dc-exs)+oextnd(b)

return dataRefined, dataCont   



n1 = 100
t = np.linspace(0,2*math.pi,n1)
y = np.sin(t)
data = quickcontmin(y)    
dc1 = data[1]
dc1 = dc1[0:n1-1]`
  • can you define your `c2` as `np.zeros((1, 2 *n - 2))*1j` ? your `c2` is a column-like array, and you're trying to put a row-like array into it... – Raf Jul 17 '17 at 20:47
  • ValueError: could not broadcast input array from shape (25) into shape (0,198) The zero array is the coefficients. That array is (25,1) –  Jul 17 '17 at 21:05
  • Can't you be more cryptic? Which array, which line, which function? We can't guess where an array or target might have shape (0, 198). – hpaulj Jul 17 '17 at 22:39
  • execfile(filename, namespace) File "C:\Python27\lib\site-packages\spyderlib\widgets\externalshell\sitecustomize.py", line 71, in execfile exec(compile(scripttext, filename, 'exec'), glob, loc) File "C:/Python27/quickmincont.py", line 141, in data = quickcontmin(y) File "C:/Python27/quickmincont.py", line 105, in quickcontmin p[:,j] = evenhp(oddev(w[:,j])) File "C:/Python27/quickmincont.py", line 56, in oddev c2[1:modes+1] = v* 1j ValueError: could not broadcast input array from shape (25) into shape (25,1) >>> It's oddev on 56 –  Jul 17 '17 at 22:54
  • You can run this you realize. –  Jul 17 '17 at 22:54
  • It's not calling anything anything else. That call is to random matrix is two filters are going over. It has 25 because it is the number of modes. which above there at the top. it is defined. –  Jul 17 '17 at 22:56
  • which is w the matrix w. on line 102 in the for loop with being initialized with the zero matrix p. there are two functions evenhp and oddev over w[:,j] there is another loop beneath it. I just want the the shape to work. –  Jul 17 '17 at 23:00

1 Answers1

0

Replacing c2[1:modes+1] = v* 1j by c2[1:modes+1, 0] = v* 1j should fix that specific error. More consistent would be to replace:

v = np.array(xv[:])
v1 = v[:-1]
v1 = v[::-1]

by

v = xv
v1 = v[:-1]

v is already a column vector so you don't need to transform it into a 1d vector when you later need a column vector.

Stephan
  • 2,028
  • 16
  • 19
  • At the time I was testing to cast it as as a complex array. However it is only holding the imaginary parts. It drops the real parts. –  Jul 17 '17 at 23:06
  • I didn't mean the *1j but why you used `(2*n - 2, 1)` instead of `2*n - 2` as argument for `np.zeros`. – Stephan Jul 17 '17 at 23:12
  • it is a column vector because it is solving a system of equations. –  Jul 17 '17 at 23:16
  • `xv` is already a column vector. Why do you transform it into a 1d vector if you later need it as column vector? – Stephan Jul 17 '17 at 23:27
  • both sides now are dimensionally the same however p[:,j] = evenhp(oddev(w[:,j])) File "C:/Python27/quickmincont.py", line 57, in oddev c2[-1 - modes + 1:-1,0] = -v1* 1j ValueError: could not broadcast input array from shape (25) into shape (24) I think when I initialized the loop it is taking that from the loop perhaps. I am missing. Or I just took something off. I think I'll slowly figure it out. –  Jul 17 '17 at 23:31
  • At least the dimensionality is correct now (even though the second option of my answer is probably the cleaner solution). The remaining error means that there is probably a `-1` to much. I would recommend you to rethink your algorithm and especially the corresponding array boundaries such that you end up with a simpler code. – Stephan Jul 17 '17 at 23:41
  • Oh there are a number of other errors in here, that I have been correcting. The boundaries towards the bottom. Those wouldn't even work by the indexing I believe. They should work now, it was in matlab and it indexes differently. –  Jul 17 '17 at 23:53