def calc_similarity_real(particle_image, projection_ctf_applied_2d, mask=None): """ Calculates the similarity based on projections :param particle_image: Experimental 2D particle image :param projection_ctf_applied_2d: :param mask: Mask for ccc :return: cross correlation coefficient """ ccc = sp_statistics.ccc(particle_image, projection_ctf_applied_2d, mask) return ccc
def compare_projs(reconfile,
classavgstack,
inputanglesdoc,
outdir,
interpolation_method=1,
log=None,
verbose=False):
"""
Make comparison stack between class averages (even-numbered (starts from 0)) and re-projections (odd-numbered).
Arguments:
reconfile : Input volume from which to generate re-projections
classavgstack ; Input image stack
inputanglesdoc : Input Euler angles doc
outdir ; Output directory
interpolation_method : Interpolation method: nearest neighbor (nn, 0), trilinear (1, default), gridding (-1)
log : Logger object
verbose : (boolean) Whether to write additional information to screen
Returns:
compstack : Stack of comparisons between input image stack (even-numbered (starts from 0)) and input volume (odd-numbered)
"""
recondata = EMAN2.EMData(reconfile)
nx = recondata.get_xsize()
# Resample reference
reconprep = prep_vol(recondata,
npad=2,
interpolation_method=interpolation_method)
ccclist = []
# Here you need actual radius to compute proper ccc's, but if you do, you have to deal with translations, PAP
mask = model_circle(nx // 2 - 2, nx, nx)
mask.write_image(os.path.join(outdir, 'maskalign.hdf'))
compstack = os.path.join(outdir, 'comp-proj-reproj.hdf')
# Number of images may have changed
nimg1 = EMAN2.EMUtil.get_image_count(classavgstack)
angleslist = read_text_row(inputanglesdoc)
for imgnum in range(nimg1):
# Get class average
classimg = get_im(classavgstack, imgnum)
# Compute re-projection
prjimg = prgl(reconprep,
angleslist[imgnum],
interpolation_method=1,
return_real=False)
# Calculate 1D power spectra
rops_dst = rops_table(classimg * mask)
rops_src = rops_table(prjimg)
# Set power spectrum of reprojection to the data.
# Since data has an envelope, it would make more sense to set data to reconstruction,
# but to do it one would have to know the actual resolution of the data.
# you can check sxprocess.py --adjpw to see how this is done properly PAP
table = [0.0] * len(rops_dst) # initialize table
for j in range(len(rops_dst)):
table[j] = sqrt(rops_dst[j] / rops_src[j])
prjimg = fft(filt_table(
prjimg,
table)) # match FFT amplitudes of re-projection and class average
cccoeff = ccc(prjimg, classimg, mask)
#print imgnum, cccoeff
classimg.set_attr_dict({'cross-corr': cccoeff})
prjimg.set_attr_dict({'cross-corr': cccoeff})
montagestack = []
montagestack.append(prjimg)
montagestack.append(classimg)
comparison_pair = montage2(montagestack, ncol=2, marginwidth=1)
comparison_pair.write_image(compstack, imgnum)
ccclist.append(cccoeff)
del angleslist
meanccc = sum(ccclist) / nimg1
print_log_msg("Average CCC is %s\n" % meanccc, log, verbose)
nimg2 = EMAN2.EMUtil.get_image_count(compstack)
for imgnum in range(nimg2): # xrange will be deprecated in Python3
prjimg = get_im(compstack, imgnum)
meanccc1 = prjimg.get_attr_default('mean-cross-corr', -1.0)
prjimg.set_attr_dict({'mean-cross-corr': meanccc})
write_header(compstack, prjimg, imgnum)
return compstack
def main():
progname = os.path.basename(sys.argv[0])
usage = progname + """ [options] <inputfile> <outputfile>
Forms chains of 2D images based on their similarities.
Functionality:
Order a 2-D stack of image based on pair-wise similarity (computed as a cross-correlation coefficent).
Options 1-3 require image stack to be aligned. The program will apply orientation parameters if present in headers.
The ways to use the program:
1. Use option initial to specify which image will be used as an initial seed to form the chain.
sp_chains.py input_stack.hdf output_stack.hdf --initial=23 --radius=25
2. If options initial is omitted, the program will determine which image best serves as initial seed to form the chain
sp_chains.py input_stack.hdf output_stack.hdf --radius=25
3. Use option circular to form a circular chain.
sp_chains.py input_stack.hdf output_stack.hdf --circular--radius=25
4. New circular code based on pairwise alignments
sp_chains.py aclf.hdf chain.hdf circle.hdf --align --radius=25 --xr=2 --pairwiseccc=lcc.txt
5. Circular ordering based on pairwise alignments
sp_chains.py vols.hdf chain.hdf mask.hdf --dd --radius=25
"""
parser = OptionParser(usage, version=SPARXVERSION)
parser.add_option(
"--dd",
action="store_true",
help="Circular ordering without adjustment of orientations",
default=False)
parser.add_option(
"--circular",
action="store_true",
help=
"Select circular ordering (first image has to be similar to the last)",
default=False)
parser.add_option(
"--align",
action="store_true",
help=
"Compute all pairwise alignments and from the table of image similarities find the best chain",
default=False)
parser.add_option(
"--initial",
type="int",
default=-1,
help=
"Specifies which image will be used as an initial seed to form the chain. (default = 0, means the first image)"
)
parser.add_option(
"--radius",
type="int",
default=-1,
help="Radius of a circular mask for similarity based ordering")
# import params for 2D alignment
parser.add_option(
"--ou",
type="int",
default=-1,
help=
"outer radius for 2D alignment < nx/2-1 (set to the radius of the particle)"
)
parser.add_option(
"--xr",
type="int",
default=0,
help="range for translation search in x direction, search is +/xr (0)")
parser.add_option(
"--yr",
type="int",
default=0,
help="range for translation search in y direction, search is +/yr (0)")
# parser.add_option("--nomirror", action="store_true", default=False, help="Disable checking mirror orientations of images (default False)")
parser.add_option("--pairwiseccc",
type="string",
default=" ",
help="Input/output pairwise ccc file")
(options, args) = parser.parse_args()
sp_global_def.BATCH = True
if options.dd:
nargs = len(args)
if nargs != 3:
ERROR("Must provide name of input and two output files!")
return
stack = args[0]
new_stack = args[1]
from sp_utilities import model_circle
from sp_statistics import ccc
from sp_statistics import mono
lend = EMUtil.get_image_count(stack)
lccc = [None] * (old_div(lend * (lend - 1), 2))
for i in range(lend - 1):
v1 = get_im(stack, i)
if (i == 0 and nargs == 2):
nx = v1.get_xsize()
ny = v1.get_ysize()
nz = v1.get_ysize()
if options.ou < 1:
radius = old_div(nx, 2) - 2
else:
radius = options.ou
mask = model_circle(radius, nx, ny, nz)
else:
mask = get_im(args[2])
for j in range(i + 1, lend):
lccc[mono(i, j)] = [ccc(v1, get_im(stack, j), mask), 0]
order = tsp(lccc)
if (len(order) != lend):
ERROR("Problem with data length")
return
sxprint("Total sum of cccs :", TotalDistance(order, lccc))
sxprint("ordering :", order)
for i in range(lend):
get_im(stack, order[i]).write_image(new_stack, i)
elif options.align:
nargs = len(args)
if nargs != 3:
ERROR("Must provide name of input and two output files!")
return
from sp_utilities import get_params2D, model_circle
from sp_fundamentals import rot_shift2D
from sp_statistics import ccc
from time import time
from sp_alignment import align2d, align2d_scf
stack = args[0]
new_stack = args[1]
d = EMData.read_images(stack)
if (len(d) < 6):
ERROR(
"Chains requires at least six images in the input stack to be executed"
)
return
"""
# will align anyway
try:
ttt = d[0].get_attr('xform.params2d')
for i in xrange(len(d)):
alpha, sx, sy, mirror, scale = get_params2D(d[i])
d[i] = rot_shift2D(d[i], alpha, sx, sy, mirror)
except:
pass
"""
nx = d[0].get_xsize()
ny = d[0].get_ysize()
if options.ou < 1:
radius = old_div(nx, 2) - 2
else:
radius = options.ou
mask = model_circle(radius, nx, ny)
if (options.xr < 0):
xrng = 0
else:
xrng = options.xr
if (options.yr < 0):
yrng = xrng
else:
yrng = options.yr
initial = max(options.initial, 0)
from sp_statistics import mono
lend = len(d)
lccc = [None] * (old_div(lend * (lend - 1), 2))
from sp_utilities import read_text_row
if options.pairwiseccc == " " or not os.path.exists(
options.pairwiseccc):
st = time()
for i in range(lend - 1):
for j in range(i + 1, lend):
# j>i meaning mono entry (i,j) or (j,i) indicates T i->j (from smaller index to larger)
# alpha, sx, sy, mir, peak = align2d(d[i],d[j], xrng, yrng, step=options.ts, first_ring=options.ir, last_ring=radius, mode = "F")
alpha, sx, sy, mir, peak = align2d_scf(d[i],
d[j],
xrng,
yrng,
ou=radius)
lccc[mono(i, j)] = [
ccc(d[j], rot_shift2D(d[i], alpha, sx, sy, mir, 1.0),
mask), alpha, sx, sy, mir
]
# print " %4d %10.1f"%(i,time()-st)
if ((not os.path.exists(options.pairwiseccc))
and (options.pairwiseccc != " ")):
write_text_row([[initial, 0, 0, 0, 0]] + lccc,
options.pairwiseccc)
elif (os.path.exists(options.pairwiseccc)):
lccc = read_text_row(options.pairwiseccc)
initial = int(lccc[0][0] + 0.1)
del lccc[0]
for i in range(len(lccc)):
T = Transform({
"type": "2D",
"alpha": lccc[i][1],
"tx": lccc[i][2],
"ty": lccc[i][3],
"mirror": int(lccc[i][4] + 0.1)
})
lccc[i] = [lccc[i][0], T]
tdummy = Transform({"type": "2D"})
maxsum = -1.023
for m in range(0, lend): # initial, initial+1):
indc = list(range(lend))
lsnake = [[m, tdummy, 0.0]]
del indc[m]
lsum = 0.0
while len(indc) > 1:
maxcit = -111.
for i in range(len(indc)):
cuc = lccc[mono(indc[i], lsnake[-1][0])][0]
if cuc > maxcit:
maxcit = cuc
qi = indc[i]
# Here we need transformation from the current to the previous,
# meaning indc[i] -> lsnake[-1][0]
T = lccc[mono(indc[i], lsnake[-1][0])][1]
# If direction is from larger to smaller index, the transformation has to be inverted
if (indc[i] > lsnake[-1][0]): T = T.inverse()
lsnake.append([qi, T, maxcit])
lsum += maxcit
del indc[indc.index(qi)]
T = lccc[mono(indc[-1], lsnake[-1][0])][1]
if (indc[-1] > lsnake[-1][0]): T = T.inverse()
lsnake.append(
[indc[-1], T, lccc[mono(indc[-1], lsnake[-1][0])][0]])
sxprint(" initial image and lsum ", m, lsum)
# print lsnake
if (lsum > maxsum):
maxsum = lsum
init = m
snake = [lsnake[i] for i in range(lend)]
sxprint(" Initial image selected : ", init, maxsum, " ",
TotalDistance([snake[m][0] for m in range(lend)], lccc))
# for q in snake: print q
from copy import deepcopy
trans = deepcopy([snake[i][1] for i in range(len(snake))])
sxprint([snake[i][0] for i in range(len(snake))])
"""
for m in xrange(lend):
prms = trans[m].get_params("2D")
print " %3d %7.1f %7.1f %7.1f %2d %6.2f"%(snake[m][0], prms["alpha"], prms["tx"], prms["ty"], prms["mirror"], snake[m][2])
"""
for k in range(lend - 2, 0, -1):
T = snake[k][1]
for i in range(k + 1, lend):
trans[i] = T * trans[i]
# To add - apply all transformations and do the overall centering.
for m in range(lend):
prms = trans[m].get_params("2D")
# print(" %3d %7.1f %7.1f %7.1f %2d %6.2f"%(snake[m][0], prms["alpha"], prms["tx"], prms["ty"], prms["mirror"], snake[m][2]) )
# rot_shift2D(d[snake[m][0]], prms["alpha"], prms["tx"], prms["ty"], prms["mirror"]).write_image(new_stack, m)
rot_shift2D(d[snake[m][0]], prms["alpha"], 0.0, 0.0,
prms["mirror"]).write_image(new_stack, m)
order = tsp(lccc)
if (len(order) != lend):
ERROR("Problem with data length")
return
sxprint(TotalDistance(order, lccc))
sxprint(order)
ibeg = order.index(init)
order = [order[(i + ibeg) % lend] for i in range(lend)]
sxprint(TotalDistance(order, lccc))
sxprint(order)
snake = [tdummy]
for i in range(1, lend):
# Here we need transformation from the current to the previous,
# meaning order[i] -> order[i-1]]
T = lccc[mono(order[i], order[i - 1])][1]
# If direction is from larger to smaller index, the transformation has to be inverted
if (order[i] > order[i - 1]): T = T.inverse()
snake.append(T)
assert (len(snake) == lend)
from copy import deepcopy
trans = deepcopy(snake)
for k in range(lend - 2, 0, -1):
T = snake[k]
for i in range(k + 1, lend):
trans[i] = T * trans[i]
# Try to smooth the angles - complicated, I am afraid one would have to use angles forward and backwards
# and find their average??
# In addition, one would have to recenter them
"""
trms = []
for m in xrange(lend):
prms = trans[m].get_params("2D")
trms.append([prms["alpha"], prms["mirror"]])
for i in xrange(3):
for m in xrange(lend):
mb = (m-1)%lend
me = (m+1)%lend
# angles order mb,m,me
# calculate predicted angles mb->m
"""
best_params = []
for m in range(lend):
prms = trans[m].get_params("2D")
# rot_shift2D(d[order[m]], prms["alpha"], prms["tx"], prms["ty"], prms["mirror"]).write_image("metro.hdf", m)
rot_shift2D(d[order[m]], prms["alpha"], 0.0, 0.0,
prms["mirror"]).write_image(args[2], m)
best_params.append(
[m, order[m], prms["alpha"], 0.0, 0.0, prms["mirror"]])
# Write alignment parameters
outdir = os.path.dirname(args[2])
aligndoc = os.path.join(outdir, "chains_params.txt")
write_text_row(best_params, aligndoc)
"""
# This was an effort to get number of loops, inconclusive, to say the least
from numpy import outer, zeros, float32, sqrt
lend = len(d)
cor = zeros(lend,float32)
cor = outer(cor, cor)
for i in xrange(lend): cor[i][i] = 1.0
for i in xrange(lend-1):
for j in xrange(i+1, lend):
cor[i,j] = lccc[mono(i,j)][0]
cor[j,i] = cor[i,j]
lmbd, eigvec = pca(cor)
from sp_utilities import write_text_file
nvec=20
print [lmbd[j] for j in xrange(nvec)]
print " G"
mm = [-1]*lend
for i in xrange(lend): # row
mi = -1.0e23
for j in xrange(nvec):
qt = eigvec[j][i]
if(abs(qt)>mi):
mi = abs(qt)
mm[i] = j
for j in xrange(nvec):
qt = eigvec[j][i]
print round(qt,3), # eigenvector
print mm[i]
print
for j in xrange(nvec):
qt = []
for i in xrange(lend):
if(mm[i] == j): qt.append(i)
if(len(qt)>0): write_text_file(qt,"loop%02d.txt"%j)
"""
"""
print [lmbd[j] for j in xrange(nvec)]
print " B"
mm = [-1]*lend
for i in xrange(lend): # row
mi = -1.0e23
for j in xrange(nvec):
qt = eigvec[j][i]/sqrt(lmbd[j])
if(abs(qt)>mi):
mi = abs(qt)
mm[i] = j
for j in xrange(nvec):
qt = eigvec[j][i]/sqrt(lmbd[j])
print round(qt,3), # eigenvector
print mm[i]
print
"""
"""
lend=3
cor = zeros(lend,float32)
cor = outer(cor, cor)
cor[0][0] =136.77
cor[0][1] = 79.15
cor[0][2] = 37.13
cor[1][0] = 79.15
cor[2][0] = 37.13
cor[1][1] = 50.04
cor[1][2] = 21.65
cor[2][1] = 21.65
cor[2][2] = 13.26
lmbd, eigvec = pca(cor)
print lmbd
print eigvec
for i in xrange(lend): # row
for j in xrange(lend): print eigvec[j][i], # eigenvector
print
print " B"
for i in xrange(lend): # row
for j in xrange(lend): print eigvec[j][i]/sqrt(lmbd[j]), # eigenvector
print
print " G"
for i in xrange(lend): # row
for j in xrange(lend): print eigvec[j][i]*sqrt(lmbd[j]), # eigenvector
print
"""
else:
nargs = len(args)
if nargs != 2:
ERROR("Must provide name of input and output file!")
return
from sp_utilities import get_params2D, model_circle
from sp_fundamentals import rot_shift2D
from sp_statistics import ccc
from time import time
from sp_alignment import align2d
stack = args[0]
new_stack = args[1]
d = EMData.read_images(stack)
try:
sxprint("Using 2D alignment parameters from header.")
ttt = d[0].get_attr('xform.params2d')
for i in range(len(d)):
alpha, sx, sy, mirror, scale = get_params2D(d[i])
d[i] = rot_shift2D(d[i], alpha, sx, sy, mirror)
except:
pass
nx = d[0].get_xsize()
ny = d[0].get_ysize()
if options.radius < 1:
radius = old_div(nx, 2) - 2
else:
radius = options.radius
mask = model_circle(radius, nx, ny)
init = options.initial
if init > -1:
sxprint(" initial image: %d" % init)
temp = d[init].copy()
temp.write_image(new_stack, 0)
del d[init]
k = 1
lsum = 0.0
while len(d) > 1:
maxcit = -111.
for i in range(len(d)):
cuc = ccc(d[i], temp, mask)
if cuc > maxcit:
maxcit = cuc
qi = i
# sxprint k, maxcit
lsum += maxcit
temp = d[qi].copy()
del d[qi]
temp.write_image(new_stack, k)
k += 1
sxprint(lsum)
d[0].write_image(new_stack, k)
else:
if options.circular:
sxprint("Using options.circular, no alignment")
# figure the "best circular" starting image
maxsum = -1.023
for m in range(len(d)):
indc = list(range(len(d)))
lsnake = [-1] * (len(d) + 1)
lsnake[0] = m
lsnake[-1] = m
del indc[m]
temp = d[m].copy()
lsum = 0.0
direction = +1
k = 1
while len(indc) > 1:
maxcit = -111.
for i in range(len(indc)):
cuc = ccc(d[indc[i]], temp, mask)
if cuc > maxcit:
maxcit = cuc
qi = indc[i]
lsnake[k] = qi
lsum += maxcit
del indc[indc.index(qi)]
direction = -direction
for i in range(1, len(d)):
if (direction > 0):
if (lsnake[i] == -1):
temp = d[lsnake[i - 1]].copy()
# print " forw ",lsnake[i-1]
k = i
break
else:
if (lsnake[len(d) - i] == -1):
temp = d[lsnake[len(d) - i + 1]].copy()
# print " back ",lsnake[len(d) - i +1]
k = len(d) - i
break
lsnake[lsnake.index(-1)] = indc[-1]
# print " initial image and lsum ",m,lsum
# print lsnake
if (lsum > maxsum):
maxsum = lsum
init = m
snake = [lsnake[i] for i in range(len(d))]
sxprint(" Initial image selected : ", init, maxsum)
sxprint(lsnake)
for m in range(len(d)):
d[snake[m]].write_image(new_stack, m)
else:
# figure the "best" starting image
sxprint("Straight chain, no alignment")
maxsum = -1.023
for m in range(len(d)):
indc = list(range(len(d)))
lsnake = [m]
del indc[m]
temp = d[m].copy()
lsum = 0.0
while len(indc) > 1:
maxcit = -111.
for i in range(len(indc)):
cuc = ccc(d[indc[i]], temp, mask)
if cuc > maxcit:
maxcit = cuc
qi = indc[i]
lsnake.append(qi)
lsum += maxcit
temp = d[qi].copy()
del indc[indc.index(qi)]
lsnake.append(indc[-1])
# sxprint " initial image and lsum ",m,lsum
# sxprint lsnake
if (lsum > maxsum):
maxsum = lsum
init = m
snake = [lsnake[i] for i in range(len(d))]
sxprint(" Initial image selected : ", init, maxsum)
sxprint(lsnake)
for m in range(len(d)):
d[snake[m]].write_image(new_stack, m)
