def shiftali_MPI(stack,
maskfile=None,
maxit=100,
CTF=False,
snr=1.0,
Fourvar=False,
search_rng=-1,
oneDx=False,
search_rng_y=-1):
from applications import MPI_start_end
from utilities import model_circle, model_blank, get_image, peak_search, get_im
from utilities import reduce_EMData_to_root, bcast_EMData_to_all, send_attr_dict, file_type, bcast_number_to_all, bcast_list_to_all
from pap_statistics import varf2d_MPI
from fundamentals import fft, ccf, rot_shift3D, rot_shift2D
from utilities import get_params2D, set_params2D
from utilities import print_msg, print_begin_msg, print_end_msg
import os
import sys
from mpi import mpi_init, mpi_comm_size, mpi_comm_rank, MPI_COMM_WORLD
from mpi import mpi_reduce, mpi_bcast, mpi_barrier, mpi_gatherv
from mpi import MPI_SUM, MPI_FLOAT, MPI_INT
from EMAN2 import Processor
from time import time
number_of_proc = mpi_comm_size(MPI_COMM_WORLD)
myid = mpi_comm_rank(MPI_COMM_WORLD)
main_node = 0
ftp = file_type(stack)
if myid == main_node:
print_begin_msg("shiftali_MPI")
max_iter = int(maxit)
if myid == main_node:
if ftp == "bdb":
from EMAN2db import db_open_dict
dummy = db_open_dict(stack, True)
nima = EMUtil.get_image_count(stack)
else:
nima = 0
nima = bcast_number_to_all(nima, source_node=main_node)
list_of_particles = list(range(nima))
image_start, image_end = MPI_start_end(nima, number_of_proc, myid)
list_of_particles = list_of_particles[image_start:image_end]
# read nx and ctf_app (if CTF) and broadcast to all nodes
if myid == main_node:
ima = EMData()
ima.read_image(stack, list_of_particles[0], True)
nx = ima.get_xsize()
ny = ima.get_ysize()
if CTF: ctf_app = ima.get_attr_default('ctf_applied', 2)
del ima
else:
nx = 0
ny = 0
if CTF: ctf_app = 0
nx = bcast_number_to_all(nx, source_node=main_node)
ny = bcast_number_to_all(ny, source_node=main_node)
if CTF:
ctf_app = bcast_number_to_all(ctf_app, source_node=main_node)
if ctf_app > 0:
ERROR("data cannot be ctf-applied", "shiftali_MPI", 1, myid)
if maskfile == None:
mrad = min(nx, ny)
mask = model_circle(mrad // 2 - 2, nx, ny)
else:
mask = get_im(maskfile)
if CTF:
from filter import filt_ctf
from morphology import ctf_img
ctf_abs_sum = EMData(nx, ny, 1, False)
ctf_2_sum = EMData(nx, ny, 1, False)
else:
ctf_2_sum = None
from global_def import CACHE_DISABLE
if CACHE_DISABLE:
data = EMData.read_images(stack, list_of_particles)
else:
for i in range(number_of_proc):
if myid == i:
data = EMData.read_images(stack, list_of_particles)
if ftp == "bdb": mpi_barrier(MPI_COMM_WORLD)
for im in range(len(data)):
data[im].set_attr('ID', list_of_particles[im])
st = Util.infomask(data[im], mask, False)
data[im] -= st[0]
if CTF:
ctf_params = data[im].get_attr("ctf")
ctfimg = ctf_img(nx, ctf_params, ny=ny)
Util.add_img2(ctf_2_sum, ctfimg)
Util.add_img_abs(ctf_abs_sum, ctfimg)
if CTF:
reduce_EMData_to_root(ctf_2_sum, myid, main_node)
reduce_EMData_to_root(ctf_abs_sum, myid, main_node)
else:
ctf_2_sum = None
if CTF:
if myid != main_node:
del ctf_2_sum
del ctf_abs_sum
else:
temp = EMData(nx, ny, 1, False)
for i in range(0, nx, 2):
for j in range(ny):
temp.set_value_at(i, j, snr)
Util.add_img(ctf_2_sum, temp)
del temp
total_iter = 0
# apply initial xform.align2d parameters stored in header
init_params = []
for im in range(len(data)):
t = data[im].get_attr('xform.align2d')
init_params.append(t)
p = t.get_params("2d")
data[im] = rot_shift2D(data[im],
p['alpha'],
sx=p['tx'],
sy=p['ty'],
mirror=p['mirror'],
scale=p['scale'])
# fourier transform all images, and apply ctf if CTF
for im in range(len(data)):
if CTF:
ctf_params = data[im].get_attr("ctf")
data[im] = filt_ctf(fft(data[im]), ctf_params)
else:
data[im] = fft(data[im])
sx_sum = 0
sy_sum = 0
sx_sum_total = 0
sy_sum_total = 0
shift_x = [0.0] * len(data)
shift_y = [0.0] * len(data)
ishift_x = [0.0] * len(data)
ishift_y = [0.0] * len(data)
for Iter in range(max_iter):
if myid == main_node:
start_time = time()
print_msg("Iteration #%4d\n" % (total_iter))
total_iter += 1
avg = EMData(nx, ny, 1, False)
for im in data:
Util.add_img(avg, im)
reduce_EMData_to_root(avg, myid, main_node)
if myid == main_node:
if CTF:
tavg = Util.divn_filter(avg, ctf_2_sum)
else:
tavg = Util.mult_scalar(avg, 1.0 / float(nima))
else:
tavg = EMData(nx, ny, 1, False)
if Fourvar:
bcast_EMData_to_all(tavg, myid, main_node)
vav, rvar = varf2d_MPI(myid, data, tavg, mask, "a", CTF)
if myid == main_node:
if Fourvar:
tavg = fft(Util.divn_img(fft(tavg), vav))
vav_r = Util.pack_complex_to_real(vav)
# normalize and mask tavg in real space
tavg = fft(tavg)
stat = Util.infomask(tavg, mask, False)
tavg -= stat[0]
Util.mul_img(tavg, mask)
# For testing purposes: shift tavg to some random place and see if the centering is still correct
#tavg = rot_shift3D(tavg,sx=3,sy=-4)
tavg = fft(tavg)
if Fourvar: del vav
bcast_EMData_to_all(tavg, myid, main_node)
sx_sum = 0
sy_sum = 0
if search_rng > 0: nwx = 2 * search_rng + 1
else: nwx = nx
if search_rng_y > 0: nwy = 2 * search_rng_y + 1
else: nwy = ny
not_zero = 0
for im in range(len(data)):
if oneDx:
ctx = Util.window(ccf(data[im], tavg), nwx, 1)
p1 = peak_search(ctx)
p1_x = -int(p1[0][3])
ishift_x[im] = p1_x
sx_sum += p1_x
else:
p1 = peak_search(Util.window(ccf(data[im], tavg), nwx, nwy))
p1_x = -int(p1[0][4])
p1_y = -int(p1[0][5])
ishift_x[im] = p1_x
ishift_y[im] = p1_y
sx_sum += p1_x
sy_sum += p1_y
if not_zero == 0:
if (not (ishift_x[im] == 0.0)) or (not (ishift_y[im] == 0.0)):
not_zero = 1
sx_sum = mpi_reduce(sx_sum, 1, MPI_INT, MPI_SUM, main_node,
MPI_COMM_WORLD)
if not oneDx:
sy_sum = mpi_reduce(sy_sum, 1, MPI_INT, MPI_SUM, main_node,
MPI_COMM_WORLD)
if myid == main_node:
sx_sum_total = int(sx_sum[0])
if not oneDx:
sy_sum_total = int(sy_sum[0])
else:
sx_sum_total = 0
sy_sum_total = 0
sx_sum_total = bcast_number_to_all(sx_sum_total, source_node=main_node)
if not oneDx:
sy_sum_total = bcast_number_to_all(sy_sum_total,
source_node=main_node)
sx_ave = round(float(sx_sum_total) / nima)
sy_ave = round(float(sy_sum_total) / nima)
for im in range(len(data)):
p1_x = ishift_x[im] - sx_ave
p1_y = ishift_y[im] - sy_ave
params2 = {
"filter_type": Processor.fourier_filter_types.SHIFT,
"x_shift": p1_x,
"y_shift": p1_y,
"z_shift": 0.0
}
data[im] = Processor.EMFourierFilter(data[im], params2)
shift_x[im] += p1_x
shift_y[im] += p1_y
# stop if all shifts are zero
not_zero = mpi_reduce(not_zero, 1, MPI_INT, MPI_SUM, main_node,
MPI_COMM_WORLD)
if myid == main_node:
not_zero_all = int(not_zero[0])
else:
not_zero_all = 0
not_zero_all = bcast_number_to_all(not_zero_all, source_node=main_node)
if myid == main_node:
print_msg("Time of iteration = %12.2f\n" % (time() - start_time))
start_time = time()
if not_zero_all == 0: break
#for im in xrange(len(data)): data[im] = fft(data[im]) This should not be required as only header information is used
# combine shifts found with the original parameters
for im in range(len(data)):
t0 = init_params[im]
t1 = Transform()
t1.set_params({
"type": "2D",
"alpha": 0,
"scale": t0.get_scale(),
"mirror": 0,
"tx": shift_x[im],
"ty": shift_y[im]
})
# combine t0 and t1
tt = t1 * t0
data[im].set_attr("xform.align2d", tt)
# write out headers and STOP, under MPI writing has to be done sequentially
mpi_barrier(MPI_COMM_WORLD)
par_str = ["xform.align2d", "ID"]
if myid == main_node:
from utilities import file_type
if (file_type(stack) == "bdb"):
from utilities import recv_attr_dict_bdb
recv_attr_dict_bdb(main_node, stack, data, par_str, image_start,
image_end, number_of_proc)
else:
from utilities import recv_attr_dict
recv_attr_dict(main_node, stack, data, par_str, image_start,
image_end, number_of_proc)
else:
send_attr_dict(main_node, data, par_str, image_start, image_end)
if myid == main_node: print_end_msg("shiftali_MPI")
def shiftali_MPI(stack, maskfile=None, maxit=100, CTF=False, snr=1.0, Fourvar=False, search_rng=-1, oneDx=False, search_rng_y=-1):
from applications import MPI_start_end
from utilities import model_circle, model_blank, get_image, peak_search, get_im
from utilities import reduce_EMData_to_root, bcast_EMData_to_all, send_attr_dict, file_type, bcast_number_to_all, bcast_list_to_all
from statistics import varf2d_MPI
from fundamentals import fft, ccf, rot_shift3D, rot_shift2D
from utilities import get_params2D, set_params2D
from utilities import print_msg, print_begin_msg, print_end_msg
import os
import sys
from mpi import mpi_init, mpi_comm_size, mpi_comm_rank, MPI_COMM_WORLD
from mpi import mpi_reduce, mpi_bcast, mpi_barrier, mpi_gatherv
from mpi import MPI_SUM, MPI_FLOAT, MPI_INT
from EMAN2 import Processor
from time import time
number_of_proc = mpi_comm_size(MPI_COMM_WORLD)
myid = mpi_comm_rank(MPI_COMM_WORLD)
main_node = 0
ftp = file_type(stack)
if myid == main_node:
print_begin_msg("shiftali_MPI")
max_iter=int(maxit)
if myid == main_node:
if ftp == "bdb":
from EMAN2db import db_open_dict
dummy = db_open_dict(stack, True)
nima = EMUtil.get_image_count(stack)
else:
nima = 0
nima = bcast_number_to_all(nima, source_node = main_node)
list_of_particles = range(nima)
image_start, image_end = MPI_start_end(nima, number_of_proc, myid)
list_of_particles = list_of_particles[image_start: image_end]
# read nx and ctf_app (if CTF) and broadcast to all nodes
if myid == main_node:
ima = EMData()
ima.read_image(stack, list_of_particles[0], True)
nx = ima.get_xsize()
ny = ima.get_ysize()
if CTF: ctf_app = ima.get_attr_default('ctf_applied', 2)
del ima
else:
nx = 0
ny = 0
if CTF: ctf_app = 0
nx = bcast_number_to_all(nx, source_node = main_node)
ny = bcast_number_to_all(ny, source_node = main_node)
if CTF:
ctf_app = bcast_number_to_all(ctf_app, source_node = main_node)
if ctf_app > 0: ERROR("data cannot be ctf-applied", "shiftali_MPI", 1, myid)
if maskfile == None:
mrad = min(nx, ny)
mask = model_circle(mrad//2-2, nx, ny)
else:
mask = get_im(maskfile)
if CTF:
from filter import filt_ctf
from morphology import ctf_img
ctf_abs_sum = EMData(nx, ny, 1, False)
ctf_2_sum = EMData(nx, ny, 1, False)
else:
ctf_2_sum = None
from global_def import CACHE_DISABLE
if CACHE_DISABLE:
data = EMData.read_images(stack, list_of_particles)
else:
for i in xrange(number_of_proc):
if myid == i:
data = EMData.read_images(stack, list_of_particles)
if ftp == "bdb": mpi_barrier(MPI_COMM_WORLD)
for im in xrange(len(data)):
data[im].set_attr('ID', list_of_particles[im])
st = Util.infomask(data[im], mask, False)
data[im] -= st[0]
if CTF:
ctf_params = data[im].get_attr("ctf")
ctfimg = ctf_img(nx, ctf_params, ny=ny)
Util.add_img2(ctf_2_sum, ctfimg)
Util.add_img_abs(ctf_abs_sum, ctfimg)
if CTF:
reduce_EMData_to_root(ctf_2_sum, myid, main_node)
reduce_EMData_to_root(ctf_abs_sum, myid, main_node)
else: ctf_2_sum = None
if CTF:
if myid != main_node:
del ctf_2_sum
del ctf_abs_sum
else:
temp = EMData(nx, ny, 1, False)
for i in xrange(0,nx,2):
for j in xrange(ny):
temp.set_value_at(i,j,snr)
Util.add_img(ctf_2_sum, temp)
del temp
total_iter = 0
# apply initial xform.align2d parameters stored in header
init_params = []
for im in xrange(len(data)):
t = data[im].get_attr('xform.align2d')
init_params.append(t)
p = t.get_params("2d")
data[im] = rot_shift2D(data[im], p['alpha'], sx=p['tx'], sy=p['ty'], mirror=p['mirror'], scale=p['scale'])
# fourier transform all images, and apply ctf if CTF
for im in xrange(len(data)):
if CTF:
ctf_params = data[im].get_attr("ctf")
data[im] = filt_ctf(fft(data[im]), ctf_params)
else:
data[im] = fft(data[im])
sx_sum=0
sy_sum=0
sx_sum_total=0
sy_sum_total=0
shift_x = [0.0]*len(data)
shift_y = [0.0]*len(data)
ishift_x = [0.0]*len(data)
ishift_y = [0.0]*len(data)
for Iter in xrange(max_iter):
if myid == main_node:
start_time = time()
print_msg("Iteration #%4d\n"%(total_iter))
total_iter += 1
avg = EMData(nx, ny, 1, False)
for im in data: Util.add_img(avg, im)
reduce_EMData_to_root(avg, myid, main_node)
if myid == main_node:
if CTF:
tavg = Util.divn_filter(avg, ctf_2_sum)
else: tavg = Util.mult_scalar(avg, 1.0/float(nima))
else:
tavg = EMData(nx, ny, 1, False)
if Fourvar:
bcast_EMData_to_all(tavg, myid, main_node)
vav, rvar = varf2d_MPI(myid, data, tavg, mask, "a", CTF)
if myid == main_node:
if Fourvar:
tavg = fft(Util.divn_img(fft(tavg), vav))
vav_r = Util.pack_complex_to_real(vav)
# normalize and mask tavg in real space
tavg = fft(tavg)
stat = Util.infomask( tavg, mask, False )
tavg -= stat[0]
Util.mul_img(tavg, mask)
# For testing purposes: shift tavg to some random place and see if the centering is still correct
#tavg = rot_shift3D(tavg,sx=3,sy=-4)
tavg = fft(tavg)
if Fourvar: del vav
bcast_EMData_to_all(tavg, myid, main_node)
sx_sum=0
sy_sum=0
if search_rng > 0: nwx = 2*search_rng+1
else: nwx = nx
if search_rng_y > 0: nwy = 2*search_rng_y+1
else: nwy = ny
not_zero = 0
for im in xrange(len(data)):
if oneDx:
ctx = Util.window(ccf(data[im],tavg),nwx,1)
p1 = peak_search(ctx)
p1_x = -int(p1[0][3])
ishift_x[im] = p1_x
sx_sum += p1_x
else:
p1 = peak_search(Util.window(ccf(data[im],tavg), nwx,nwy))
p1_x = -int(p1[0][4])
p1_y = -int(p1[0][5])
ishift_x[im] = p1_x
ishift_y[im] = p1_y
sx_sum += p1_x
sy_sum += p1_y
if not_zero == 0:
if (not(ishift_x[im] == 0.0)) or (not(ishift_y[im] == 0.0)):
not_zero = 1
sx_sum = mpi_reduce(sx_sum, 1, MPI_INT, MPI_SUM, main_node, MPI_COMM_WORLD)
if not oneDx:
sy_sum = mpi_reduce(sy_sum, 1, MPI_INT, MPI_SUM, main_node, MPI_COMM_WORLD)
if myid == main_node:
sx_sum_total = int(sx_sum[0])
if not oneDx:
sy_sum_total = int(sy_sum[0])
else:
sx_sum_total = 0
sy_sum_total = 0
sx_sum_total = bcast_number_to_all(sx_sum_total, source_node = main_node)
if not oneDx:
sy_sum_total = bcast_number_to_all(sy_sum_total, source_node = main_node)
sx_ave = round(float(sx_sum_total)/nima)
sy_ave = round(float(sy_sum_total)/nima)
for im in xrange(len(data)):
p1_x = ishift_x[im] - sx_ave
p1_y = ishift_y[im] - sy_ave
params2 = {"filter_type" : Processor.fourier_filter_types.SHIFT, "x_shift" : p1_x, "y_shift" : p1_y, "z_shift" : 0.0}
data[im] = Processor.EMFourierFilter(data[im], params2)
shift_x[im] += p1_x
shift_y[im] += p1_y
# stop if all shifts are zero
not_zero = mpi_reduce(not_zero, 1, MPI_INT, MPI_SUM, main_node, MPI_COMM_WORLD)
if myid == main_node:
not_zero_all = int(not_zero[0])
else:
not_zero_all = 0
not_zero_all = bcast_number_to_all(not_zero_all, source_node = main_node)
if myid == main_node:
print_msg("Time of iteration = %12.2f\n"%(time()-start_time))
start_time = time()
if not_zero_all == 0: break
#for im in xrange(len(data)): data[im] = fft(data[im]) This should not be required as only header information is used
# combine shifts found with the original parameters
for im in xrange(len(data)):
t0 = init_params[im]
t1 = Transform()
t1.set_params({"type":"2D","alpha":0,"scale":t0.get_scale(),"mirror":0,"tx":shift_x[im],"ty":shift_y[im]})
# combine t0 and t1
tt = t1*t0
data[im].set_attr("xform.align2d", tt)
# write out headers and STOP, under MPI writing has to be done sequentially
mpi_barrier(MPI_COMM_WORLD)
par_str = ["xform.align2d", "ID"]
if myid == main_node:
from utilities import file_type
if(file_type(stack) == "bdb"):
from utilities import recv_attr_dict_bdb
recv_attr_dict_bdb(main_node, stack, data, par_str, image_start, image_end, number_of_proc)
else:
from utilities import recv_attr_dict
recv_attr_dict(main_node, stack, data, par_str, image_start, image_end, number_of_proc)
else: send_attr_dict(main_node, data, par_str, image_start, image_end)
if myid == main_node: print_end_msg("shiftali_MPI")