def update_patcher(self):
""" initialize and update patcher """
def need_reget_label():
""" check if need to get label from patcher """
# check if haven't got labels from patcher
if not "labels" in self.database[self.index]:
return True
# check if have got labels from patcher, but with None label file
# note: there is a case that this will lead to excess label getting,
# that is, patcher.labels contains no labels.
count = 0
for class_i,boxes in self.database[self.index]["labels"].items():
count += len(boxes)
return count == 0
# in case when patcher initialized without label file, need to reinitialize
if need_reget_label():
self.patcher = Patcher(self.database[self.index]["fname"], self.database[self.index]["lname"])
self.database[self.index]["labels"] = self.patcher.get_labels()
def __init__(self):
self.trainptchs = []
self.traincount = 1
self.testptchs = []
self.testcount = 1
sumptraintchs = []
sumptesttchs = []
orgtraindata = getData(trnTst='training')
print(orgtraindata.shape)
orgtestdata = getData(trnTst='testing')
print('orgtraindata', orgtraindata.dtype, type(orgtraindata))
Ptchr = Patcher(imsize=[orgtraindata.shape[1], orgtraindata.shape[2]],
patchsize=28,
step=int(28 / 2),
nopartials=True,
contatedges=True)
#Ptchr1=Patcher(imsize=[orgtestdata.shape[1],orgtestdata.shape[2]],patchsize=28,step=int(28/2), nopartials=True, contatedges=True)
for i in range(500):
ptchs_1 = Ptchr.im2patches(
add_Usp_nous(orgtraindata[i, :, :, :]).reshape([256, 256]))
sumptraintchs += ptchs_1
#orgtraindata[i,:,:,0] = np.reshape(add_Usp_nous(orgtraindata[i,:,:,:]).reshape([64,64]), [64,64])
for i in range(4):
ptchs_1 = Ptchr.im2patches(
add_Usp_nous(orgtestdata[i, :, :, :]).reshape([256, 256]))
sumptesttchs += ptchs_1
sumptraintchs = np.array(sumptraintchs)
sumptraintchs = sumptraintchs.reshape(-1, 784)
print(sumptraintchs.shape)
self.trainptchs = sumptraintchs
sumptesttchs = np.array(sumptesttchs)
sumptesttchs = sumptesttchs.reshape(-1, 784)
self.testptchs = sumptesttchs
for (option, enum_name, requires_parameter) in opts_long:
if first:
arg_enum.append(" %s = 1000," % (enum_name))
first = False
else:
arg_enum.append(" %s," % (enum_name))
arg_enum.append("};")
arg_enum = "\n".join(arg_enum) + "\n"
cmd_def = [" const char *short_options = \"%s\";" % (short_string)]
cmd_def.append(" struct option long_options[] = {")
for (option, enum_name, requires_parameter) in opts_long:
param = "\"%s\"," % (option)
if requires_parameter:
cmd_def.append(" { %-30s required_argument, 0, %s }," %
(param, enum_name))
else:
cmd_def.append(" { %-30s no_argument, 0, %s }," %
(param, enum_name))
cmd_def.append(" { 0 }")
cmd_def.append(" };")
cmd_def = "\n".join(cmd_def) + "\n"
patcher = Patcher("../pgmopts.c")
patcher.patch("help page", help_code)
patcher.patch("command definition enum", arg_enum)
patcher.patch("command definition", cmd_def)
patcher = Patcher("../README.md", filetype="markdown")
patcher.patch("help page", markdown_help_page)
def vaerecon(us_ksp_r2,
sensmaps,
dcprojiter,
n=10,
lat_dim=60,
patchsize=28,
contRec='',
parfact=10,
num_iter=302,
rescaled=False,
half=False,
regiter=15,
reglmb=0.1,
regtype='reg2_dc',
usemeth=1,
stepsize=1e-4,
optScale=False,
mode=[],
chunks40=False,
Melmodels='',
N4BFcorr=False,
z_multip=1.0,
directapprox=0,
vae_model='',
logdir='',
directapp=0,
gt=None):
print('xxxxxxxxxxxxxxxxxxx contRec is ' + contRec)
print('xxxxxxxxxxxxxxxxxxx parfact is ' + str(parfact))
import pickle
# set parameters
# ==============================================================================
np.random.seed(seed=1)
imsizer = us_ksp_r2.shape[0]
imrizec = us_ksp_r2.shape[1]
nsampl = 50 # 0
# make a network and a patcher to use later
# ==============================================================================
x_rec, x_inp, funop, grd0, grd_dir, sess, grd_p_x_z0, grd_p_z0, grd_q_z_x0, grd20, y_out, y_out_prec, z_std_multip, op_q_z_x, mu, std, grd_q_zpl_x_az0, op_q_zpl_x, z_pl, z = definevae(
lat_dim=lat_dim,
patchsize=patchsize,
mode=mode,
vae_model=vae_model,
batchsize=parfact * nsampl)
if directapp:
print('_____DIRECT APPROX_____')
grd0 = grd_dir
Ptchr = Patcher(imsize=[imsizer, imrizec],
patchsize=patchsize,
step=int(patchsize / 2),
nopartials=True,
contatedges=True)
nopatches = len(Ptchr.genpatchsizes)
print("KCT-INFO: there will be in total " + str(nopatches) + " patches.")
# define the necessary functions
# ==============================================================================
def FT(x):
# inp: [nx, ny]
# out: [nx, ny, ns]
return np.fft.fftshift(np.fft.fft2(
sensmaps * np.tile(x[:, :, np.newaxis], [1, 1, sensmaps.shape[2]]),
axes=(0, 1)),
axes=(0, 1))
# def tFT(x):
# # inp: [nx, ny, ns]
# # out: [nx, ny]
# tft_x = np.fft.ifft2(np.fft.ifftshift(x, axes=(0, 1)), axes=(0, 1)) * np.conjugate(sensmaps)
#
# rss = np.sqrt(np.sum(np.square(tft_x), axis=2))
#
# rss = rss / (np.sqrt(np.sum(np.square(sensmaps*np.conjugate(sensmaps)),axis=2)) + 0.00000001)
#
# return rss # root-sum-squared
def tFT(x):
# inp: [nx, ny, ns]
# out: [nx, ny]
temp = np.fft.ifft2(np.fft.ifftshift(x, axes=(0, 1)), axes=(0, 1))
return np.sum(temp * np.conjugate(sensmaps), axis=2) / (
np.sum(sensmaps * np.conjugate(sensmaps), axis=2) + 0.00000001)
def UFT(x, uspat):
# inp: [nx, ny], [nx, ny]
# out: [nx, ny, ns]
return np.tile(uspat[:, :, np.newaxis],
[1, 1, sensmaps.shape[2]]) * FT(x)
def tUFT(x, uspat):
# inp: [nx, ny], [nx, ny]
# out: [nx, ny]
tmp1 = np.tile(uspat[:, :, np.newaxis], [1, 1, sensmaps.shape[2]])
return tFT(tmp1 * x)
def dconst(us):
# inp: [nx, ny]
# out: [nx, ny]
return np.linalg.norm(UFT(us, uspat) - data)**2
def dconst_grad(us):
# inp: [nx, ny]
# out: [nx, ny]
return 2 * tUFT(UFT(us, uspat) - data, uspat)
def likelihood(us):
# inp: [parfact,ps*ps]
# out: parfact
us = np.abs(us)
funeval = funop.eval(feed_dict={
x_rec: np.tile(us, (nsampl, 1)),
z_std_multip: z_multip
}) # ,x_inp: np.tile(us,(nsampl,1))
# funeval: [500x1]
funeval = np.array(np.split(funeval, nsampl,
axis=0)) # [nsampl x parfact x 1]
return np.mean(funeval, axis=0).astype(np.float64)
def likelihood_grad(us):
# inp: [parfact, ps*ps]
# out: [parfact, ps*ps]
usc = us.copy()
usabs = np.abs(us)
grd0eval = grd0.eval(feed_dict={
x_rec: np.tile(usabs, (nsampl, 1)),
z_std_multip: z_multip
}) # ,x_inp: np.tile(usabs,(nsampl,1))
# grd0eval: [500x784]
grd0eval = np.array(np.split(grd0eval, nsampl,
axis=0)) # [nsampl x parfact x 784]
sigmaeval = y_out_prec.eval(feed_dict={
x_rec: np.tile(usabs, (nsampl, 1)),
z_std_multip: z_multip
}) # ,x_inp: np.tile(usabs,(nsampl,1))
sigmaeval = np.array(np.split(sigmaeval, nsampl,
axis=0)) # [nsampl x parfact x 784]
mueval = y_out.eval(feed_dict={
x_rec: np.tile(usabs, (nsampl, 1)),
z_std_multip: z_multip
}) # ,x_inp: np.tile(usabs,(nsampl,1))
mueval = np.array(np.split(mueval, nsampl,
axis=0)) # [nsampl x parfact x 784]
#vareval = np.std(mueval, axis=0) # V(MU(X))
#vareval = np.mean(1/sigmaeval, axis=0) # M(SIGMA)
vareval = np.std(grd0eval, axis=0) # V(SIGMA (X-MU(X)))
# grd0_var = np.std(grd0eval, axis=0)
grd0m = np.mean(grd0eval, axis=0) # [parfact,784]
#grd0m = usc / np.abs(usc) * grd0m
where_not_0 = np.where(usc > 0)
div = usc
div[where_not_0] = usc[where_not_0] / np.abs(usc)[where_not_0].astype(
'float')
grd0m = div * grd0m
var0m = vareval
return grd0m, var0m # .astype(np.float64)
def likelihood_grad_meth3(us):
# inp: [parfact, ps*ps]
# out: [parfact, ps*ps]
usc = us.copy()
usabs = np.abs(us)
mueval = mu.eval(feed_dict={x_rec: np.tile(usabs, (nsampl, 1))
}) # ,x_inp: np.tile(usabs,(nsampl,1))
# print("===============================================================")
# print("===============================================================")
# print("===============================================================")
# print(mueval)
# print("===============================================================")
# print("===============================================================")
# print("===============================================================")
# print(mueval.shape)
# print("===============================================================")
# print("===============================================================")
# print("===============================================================")
# print(len(mueval))
# print("===============================================================")
# print("===============================================================")
# print("===============================================================")
stdeval = std.eval(feed_dict={x_rec: np.tile(usabs, (nsampl, 1))
}) # ,x_inp: np.tile(usabs,(nsampl,1))
zvals = mueval + np.random.rand(mueval.shape[0],
mueval.shape[1]) * stdeval
y_outeval = y_out.eval(feed_dict={z: zvals})
y_out_preceval = y_out_prec.eval(feed_dict={z: zvals})
tmp = np.tile(usabs, (nsampl, 1)) - y_outeval
tmp = (-1) * tmp * y_out_preceval
# grd0eval: [500x784]
grd0eval = np.array(np.split(tmp, nsampl,
axis=0)) # [nsampl x parfact x 784]
grd0m = np.mean(grd0eval, axis=0) # [parfact,784]
where_not_0 = np.where(usc > 0)
div = usc
div[where_not_0] = usc[where_not_0] / np.abs(usc)[where_not_0]
grd0m = div * grd0m
return grd0m # .astype(np.float64)
def likelihood_grad_patches(ptchs):
# inp: [np, ps, ps]
# out: [np, ps, ps]
# takes set of patches as input and returns a set of their grad.s
# both grads are in the positive direction
shape_orig = ptchs.shape
ptchs = np.reshape(ptchs, [ptchs.shape[0], -1])
grds = np.zeros([
int(np.ceil(ptchs.shape[0] / parfact) * parfact),
np.prod(ptchs.shape[1:])
],
dtype=np.complex64)
grds_vars = grds.copy()
extraind = int(
np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, ((0, extraind), (0, 0)), mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
if usemeth == 1:
grds[parfact * ix:parfact * ix +
parfact, :], grds_vars[parfact * ix:parfact * ix +
parfact, :] = likelihood_grad(
ptchs[parfact *
ix:parfact * ix +
parfact, :])
else:
assert (1 == 0)
grds = grds[0:shape_orig[0], :]
grds_vars = grds_vars[0:shape_orig[0], :]
return np.reshape(grds, shape_orig), np.reshape(grds_vars, shape_orig)
def likelihood_patches(ptchs):
# inp: [np, ps, ps]
# out: 1
fvls = np.zeros([int(np.ceil(ptchs.shape[0] / parfact) * parfact)])
extraind = int(
np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, [(0, extraind), (0, 0), (0, 0)], mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
fvls[parfact * ix:parfact * ix + parfact] = likelihood(
np.reshape(ptchs[parfact * ix:parfact * ix + parfact, :, :],
[parfact, -1]))
fvls = fvls[0:ptchs.shape[0]]
return np.mean(fvls)
def full_gradient(image):
# inp: [nx*nx, 1]
# out: [nx, ny], [nx, ny]
# returns both gradients in the respective positive direction.
# i.e. must
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer, imrizec]))
ptchs = np.array(ptchs)
grd_lik, grd_lik_var = likelihood_grad_patches(ptchs)
grd_lik = (-1) * Ptchr.patches2im(grd_lik)
grd_lik_var = Ptchr.patches2im(grd_lik_var)
grd_dconst = dconst_grad(np.reshape(image, [imsizer, imrizec]))
return grd_lik + grd_dconst, grd_lik, grd_dconst, grd_lik_var
def full_funceval(image):
# inp: [nx*nx, 1]
# out: [1], [1], [1]
tmpimg = np.reshape(image, [imsizer, imrizec])
dc = dconst(tmpimg)
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer, imrizec]))
ptchs = np.array(ptchs)
lik = (-1) * likelihood_patches(np.abs(ptchs))
return lik + dc, lik, dc
def tv_proj(phs, mu=0.125, lmb=2, IT=225):
phs = fb_tv_proj(phs, mu=mu, lmb=lmb, IT=IT)
return phs
def fgrad(im):
imr_x = np.roll(im, shift=-1, axis=0)
imr_y = np.roll(im, shift=-1, axis=1)
grd_x = imr_x - im
grd_y = imr_y - im
return np.array((grd_x, grd_y))
def fdivg(im):
imr_x = np.roll(np.squeeze(im[0, :, :]), shift=1, axis=0)
imr_y = np.roll(np.squeeze(im[1, :, :]), shift=1, axis=1)
grd_x = np.squeeze(im[0, :, :]) - imr_x
grd_y = np.squeeze(im[1, :, :]) - imr_y
return grd_x + grd_y
def f_st(u, lmb):
uabs = np.squeeze(np.sqrt(np.sum(u * np.conjugate(u), axis=0)))
tmp = 1 - lmb / uabs
tmp[np.abs(tmp) < 0] = 0
uu = u * np.tile(tmp[np.newaxis, :, :], [u.shape[0], 1, 1])
return uu
def fb_tv_proj(im, u0=0, mu=0.125, lmb=1, IT=15):
sz = im.shape
us = np.zeros((2, sz[0], sz[1], IT))
us[:, :, :, 0] = u0
for it in range(IT - 1):
# grad descent step:
tmp1 = im - fdivg(us[:, :, :, it])
tmp2 = mu * fgrad(tmp1)
tmp3 = us[:, :, :, it] - tmp2
# thresholding step:
us[:, :, :, it + 1] = tmp3 - f_st(tmp3, lmb=lmb)
# endfor
return im - fdivg(us[:, :, :, it + 1])
def g_tv_eval(x):
x_re = np.fft.fftshift(np.reshape(x, (imsizer, imrizec, 1)),
axes=(0, 1))
data = tf.placeholder(tf.float64, shape=x_re.shape)
x_tv = tf.image.total_variation(data)
var_grad = tf.gradients(x_tv, [data])[0]
var_grad_val = var_grad.eval(feed_dict={data: x_re})
return np.fft.ifftshift(var_grad_val, axes=(0, 1))
def tv_norm(x):
"""Computes the total variation norm and its gradient. From jcjohnson/cnn-vis."""
x = np.fft.fftshift(np.reshape(x, (imsizer, imrizec, 1)), axes=(0, 1))
x_diff = x - np.roll(x, -1, axis=1)
y_diff = x - np.roll(x, -1, axis=0)
grad_norm2 = x_diff**2 + y_diff**2 + np.finfo(np.float32).eps
norm = np.sum(np.sqrt(grad_norm2))
dgrad_norm = 0.5 / np.sqrt(grad_norm2)
dx_diff = 2 * x_diff * dgrad_norm
dy_diff = 2 * y_diff * dgrad_norm
grad = dx_diff + dy_diff
grad[:, 1:, :] -= dx_diff[:, :-1, :]
grad[1:, :, :] -= dy_diff[:-1, :, :]
return norm, np.reshape(np.fft.ifftshift(grad, axes=(0, 1)), [-1])
# make the data
# ===============================
uspat = np.abs(us_ksp_r2) > 0
uspat = uspat[:, :, 0]
data = us_ksp_r2
trpat = np.zeros_like(uspat)
trpat[:, 120:136] = 1
# lrphase = np.angle( tUFT(data*trpat[:,:,np.newaxis],uspat) )
# lrphase = pickle.load(open('/home/ktezcan/unnecessary_stuff/lowresphase','rb'))
# truephase = pickle.load(open('/home/ktezcan/unnecessary_stuff/truephase','rb'))
# lrphase = pickle.load(open('/home/ktezcan/unnecessary_stuff/usphase','rb'))
# lrphase = pickle.load(open('/home/ktezcan/unnecessary_stuff/lrusphase','rb'))
# lrphase = pickle.load(open('/home/ktezcan/unnecessary_stuff/lrmaskphase','rb'))
# make the functions for POCS
# =====================================
numiter = num_iter
multip = 0 # 0.1
alphas = stepsize * np.ones(numiter) # np.logspace(-4,-4,numiter)
# alphas=np.ones_like(np.logspace(-4,-4,numiter))*5e-3
def feval(im):
return full_funceval(im)
def geval(im):
t1, t2, t3, t4 = full_gradient(im)
return np.reshape(t1, [-1]), np.reshape(t2, [-1]), np.reshape(
t3, [-1]), np.reshape(t4, [-1])
# initialize data
recs = np.zeros((imsizer * imrizec, numiter + 2), dtype=complex)
# recs[:,0] = np.abs(tUFT(data, uspat).flatten().copy()) #kct
recs[:, 0] = tUFT(data, uspat).flatten().copy()
#pickle.dump(recs[:, 0], open(logdir + '_rec_0', 'wb'))
n4bf = 1
# recs[:,0] = np.abs(tUFT(data, uspat).flatten().copy() )*np.exp(1j*lrphase).flatten()
phaseregvals = []
# pickle.dump(recs[:,0],open('/scratc_','wb'))
print('contRec is ' + contRec)
if contRec != '':
try:
print('KCT-INFO: reading from a previous pickle file ' + contRec)
import pickle
rr = pickle.load(open(contRec, 'rb'))
recs[:, 0] = rr[:, -1]
print('KCT-INFO: initialized to the previous recon from pickle: ' +
contRec)
except:
print('KCT-INFO: reading from a previous numpy file ' + contRec)
rr = np.load(contRec)
recs[:, 0] = rr[:, -1]
print('KCT-INFO: initialized to the previous recon from numpy: ' +
contRec)
n4biasfields = []
recsarr = []
for it in range(0, numiter - 2, 2):
alpha = alphas[it]
# first do N times magnitude prior iterations
# ===============================================
# ===============================================
recstmp = recs[:, it].copy()
ftot, f_lik, f_dc = 0, 0, 0 #feval(recstmp)
gtot, g_lik, g_dc, g_lik_var = geval(recstmp)
tvnorm, tvgrad = tv_norm(np.abs(recstmp))
lambda_lik = 0
lambda_reg = 1
recstmp_1 = recstmp - alpha * (lambda_lik * g_lik +
lambda_reg * tvgrad)
recs[:, it + 1] = recstmp_1.copy()
print("it no: " + str(it) + " f_tot= " + str(ftot) + " f_lik= " +
str(f_lik) + ' TV norm= ' + str(tvnorm) + " f_dc (1e6)= " +
str(f_dc / 1e6) + " |g_lik|= " + str(np.linalg.norm(g_lik)) +
" |g_dc|= " + str(np.linalg.norm(g_dc)) + ' |g_tv|= ' +
str(np.linalg.norm(tvgrad)))
# if it == 0:
# pickle.dump(recstmp, open(logdir + '_rec_0', 'wb'))
# pickle.dump(g_lik, open(logdir + '_rec_likgrad', 'wb'))
# pickle.dump(g_dc, open(logdir + '_rec_dcgrad', 'wb'))
# pickle.dump(tvgrad, open(logdir + '_rec_tvgrad', 'wb'))
# pickle.dump(tvgrad * g_lik_var, open(logdir + '_rec_tvmulvar', 'wb'))
# pickle.dump(g_lik_var, open(logdir + '_rec_var', 'wb'))
# exit()
# now do again a data consistency projection
# ===============================================
# ===============================================
tmp1 = UFT(np.reshape(recs[:, it + 1], [imsizer, imrizec]),
(1 - uspat))
tmp2 = UFT(np.reshape(recs[:, it + 1], [imsizer, imrizec]), (uspat))
tmp3 = data * uspat[:, :, np.newaxis]
tmp = tmp1 + multip * tmp2 + (1 - multip) * tmp3
recs[:, it + 2] = tFT(tmp).flatten()
#ftot, f_lik, f_dc = feval(recs[:, it + 2])
#print('f_dc (1e6): ' + str(f_dc / 1e6) + ' perc: ' + str(100 * f_dc / np.linalg.norm(data) ** 2))
# MSE CHECK
recon_sli = np.reshape(recs[:, it + 2], (imsizer, imrizec))
gt = np.reshape(gt, (imsizer, imrizec))
rss = np.sqrt(
np.sum(np.square(
np.abs(sensmaps * np.tile(recon_sli[:, :, np.newaxis],
[1, 1, sensmaps.shape[2]]))),
axis=-1))
nmse = np.sqrt(((np.fft.fftshift(rss) - gt)**2).mean()) / np.sqrt(
((gt)**2).mean())
print('NMSE: ', nmse)
return recs, 0, phaseregvals, n4biasfields
#!/usr/bin/env python2
# Vocoder Patch for MD380 Firmware
# Applies to version S013.020
from Patcher import Patcher
#Match all public calls.
monitormode = False
#Match all private calls.
monitormodeprivate = False
if __name__ == '__main__':
print "Creating patches from unwrapped.img."
patcher = Patcher("unwrapped.img")
# bypass vocoder copy protection on S013.020
patcher.nopout((0x8034a60))
patcher.nopout((0x8034a60 + 0x2))
patcher.nopout((0x8034a76))
patcher.nopout((0x8034a76 + 0x2))
patcher.nopout((0x8034a8c))
patcher.nopout((0x8034a8c + 0x2))
patcher.nopout((0x8034aa2))
patcher.nopout((0x8034aa2 + 0x2))
patcher.nopout((0x8034ab8))
patcher.nopout((0x8034ab8 + 0x2))
patcher.nopout((0x8034ace))
patcher.nopout((0x8034ace + 0x2))
patcher.nopout((0x8049f9a))
patcher.nopout((0x8049f9a + 0x2))
patcher.nopout((0x804a820))
def vaerecon(us_ksp_r2,
dcprojiter,
onlydciter=0,
lat_dim=60,
patchsize=28,
contRec='',
lowresmodel=False,
parfact=10,
num_iter=1,
regiter=15,
reglmb=0.05,
regtype='TV'):
print('KCT-INFO: contRec is ' + contRec)
print('KCT-INFO: parfact is ' + str(parfact))
print("............", us_ksp_r2.shape)
#print(sensmaps.shape)
# set parameters
#==============================================================================
np.random.seed(seed=1)
imsizer = us_ksp_r2.shape[0] #252#256#252
imrizec = us_ksp_r2.shape[1] #308#256#308
print(imsizer, imrizec)
nsampl = 50
#make a network and a patcher to use later
#==============================================================================
x_rec, x_inp, funop, grd0, _, _, _, _, _, _, _, _, _, _, _, _, _, _ = definevae(
lat_dim=lat_dim, patchsize=patchsize, batchsize=parfact * nsampl)
Ptchr = Patcher(imsize=[imsizer, imrizec],
patchsize=patchsize,
step=int(patchsize / 2),
nopartials=True,
contatedges=True)
nopatches = len(Ptchr.genpatchsizes)
print("KCT-INFO: there will be in total " + str(nopatches) + " patches.")
#define the necessary functions
#==============================================================================
def FT(x):
# coil expansion followed by Fourier transform
#inp: [nx, ny]
#out: [nx, ny, ns]
return np.fft.fftshift(np.fft.fft2(x[:, :, np.newaxis], axes=(0, 1)),
axes=(0, 1))
def tFT(x):
# inverse Fourier transform and coil combination
#inp: [nx, ny, ns]
#out: [nx, ny]
temp = np.fft.ifft2(np.fft.ifftshift(x, axes=(0, 1)), axes=(0, 1))
return np.sum(temp, axis=2)
def UFT(x, uspat):
# Encoding: undersampling + FT
#inp: [nx, ny], [nx, ny]
#out: [nx, ny, ns]
return uspat[:, :, np.newaxis] * FT(x)
def tUFT(x, uspat):
# transposed Encoding: inverse FT + undersampling
#inp: [nx, ny], [nx, ny]
#out: [nx, ny]
tmp1 = uspat[:, :, np.newaxis]
return tFT(tmp1 * x)
def dconst(us):
#inp: [nx, ny]
#out: [nx, ny]
return np.linalg.norm(UFT(us, uspat) - data)**2
#np.linalg.norm ji suan l2 fanshu
def dconst_grad(us):
#inp: [nx, ny]
#out: [nx, ny]
return 2 * tUFT(UFT(us, uspat) - data, uspat)
def prior(us):
#inp: [parfact,ps*ps]
#out: parfact
us = np.abs(us)
funeval = funop.eval(feed_dict={x_rec: np.tile(us, (nsampl, 1))}) #
funeval = np.array(np.split(funeval, nsampl,
axis=0)) # [nsampl x parfact x 1]
return np.mean(funeval, axis=0).astype(np.float64)
def prior_grad(us):
#inp: [parfact, ps*ps]
#out: [parfact, ps*ps]
usc = us.copy()
usabs = np.abs(us)
grd0eval = grd0.eval(feed_dict={x_rec: np.tile(usabs, (nsampl, 1))
}) # ,x_inp: np.tile(usabs,(nsampl,1))
#grd0eval: [500x784]
grd0eval = np.array(np.split(grd0eval, nsampl,
axis=0)) # [nsampl x parfact x 784]
grd0m = np.mean(grd0eval, axis=0) #[parfact,784]
print("11111111111111111111111111111111")
print(np.abs(usc))
print("22222222222222222222222222222222")
print(grd0m)
grd0m = usc / np.abs(usc) * grd0m
return grd0m #.astype(np.float64)
def prior_grad_patches(ptchs):
#inp: [np, ps, ps]
#out: [np, ps, ps]
#takes set of patches as input and returns a set of their grad.s
#both grads are in the positive direction
shape_orig = ptchs.shape
ptchs = np.reshape(ptchs, [ptchs.shape[0], -1])
grds = np.zeros([
int(np.ceil(ptchs.shape[0] / parfact) * parfact),
np.prod(ptchs.shape[1:])
],
dtype=np.complex64)
extraind = int(
np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, ((0, extraind), (0, 0)), mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
grds[parfact * ix:parfact * ix + parfact, :] = prior_grad(
ptchs[parfact * ix:parfact * ix + parfact, :])
grds = grds[0:shape_orig[0], :]
return np.reshape(grds, shape_orig)
#np.ceil xiang shang quzhengshu
def prior_patches(ptchs):
#inp: [np, ps, ps]
#out: 1
fvls = np.zeros([int(np.ceil(ptchs.shape[0] / parfact) * parfact)])
extraind = int(
np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, [(0, extraind), (0, 0), (0, 0)], mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
fvls[parfact * ix:parfact * ix + parfact] = prior(
np.reshape(ptchs[parfact * ix:parfact * ix + parfact, :, :],
[parfact, -1]))
fvls = fvls[0:ptchs.shape[0]]
return np.mean(fvls)
def full_gradient(image):
#inp: [nx*nx, 1]
#out: [nx, ny], [nx, ny]
#returns both gradients in the respective positive direction.
#i.e. must
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer, imrizec]))
ptchs = np.array(ptchs)
grd_prior = prior_grad_patches(ptchs)
grd_prior = (-1) * Ptchr.patches2im(grd_prior)
grd_dconst = dconst_grad(np.reshape(image, [imsizer, imrizec]))
return grd_prior + grd_dconst, grd_prior, grd_dconst
def full_funceval(image):
#inp: [nx*nx, 1]
#out: [1], [1], [1]
tmpimg = np.reshape(image, [imsizer, imrizec])
dc = dconst(tmpimg)
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer, imrizec]))
ptchs = np.array(ptchs)
priorval = (-1) * prior_patches(np.abs(ptchs))
return priorval + dc, priorval, dc
#define the phase regularization functions
#==============================================================================
def tv_proj(phs, mu=0.125, lmb=2, IT=225):
# Total variation based projection
phs = fb_tv_proj(phs, mu=mu, lmb=lmb, IT=IT)
return phs
def fgrad(im):
# gradient operation with 1st order finite differences
imr_x = np.roll(im, shift=-1, axis=0)
imr_y = np.roll(im, shift=-1, axis=1)
grd_x = imr_x - im
grd_y = imr_y - im
return np.array((grd_x, grd_y))
def fdivg(im):
# divergence operator with 1st order finite differences
imr_x = np.roll(np.squeeze(im[0, :, :]), shift=1, axis=0)
imr_y = np.roll(np.squeeze(im[1, :, :]), shift=1, axis=1)
grd_x = np.squeeze(im[0, :, :]) - imr_x
grd_y = np.squeeze(im[1, :, :]) - imr_y
return grd_x + grd_y
def f_st(u, lmb):
# soft thresholding
uabs = np.squeeze(np.sqrt(np.sum(u * np.conjugate(u), axis=0)))
tmp = 1 - lmb / uabs
tmp[np.abs(tmp) < 0] = 0
uu = u * np.tile(tmp[np.newaxis, :, :], [u.shape[0], 1, 1])
return uu
def fb_tv_proj(im, u0=0, mu=0.125, lmb=1, IT=15):
sz = im.shape
us = np.zeros((2, sz[0], sz[1], IT))
us[:, :, :, 0] = u0
for it in range(IT - 1):
#grad descent step:
tmp1 = im - fdivg(us[:, :, :, it])
tmp2 = mu * fgrad(tmp1)
tmp3 = us[:, :, :, it] - tmp2
#thresholding step:
us[:, :, :, it + 1] = tmp3 - f_st(tmp3, lmb=lmb)
#endfor
return im - fdivg(us[:, :, :, it + 1])
def reg2_proj(usph, niter=100, alpha=0.05):
#A smoothness based based projection. Regularization method 2 from
#"Separate Magnitude and Phase Regularization via Compressed Sensing", Feng Zhao et al, IEEE TMI, 2012
usph = usph + np.pi
ims = np.zeros((imsizer, imrizec, niter))
ims[:, :, 0] = usph.copy()
for ix in range(niter - 1):
ims[:, :, ix + 1] = ims[:, :, ix] - 2 * alpha * np.real(
1j * np.exp(-1j * ims[:, :, ix]) *
fdivg(fgrad(np.exp(1j * ims[:, :, ix]))))
return ims[:, :, -1] - np.pi
#make the data
#===============================
uspat = np.abs(us_ksp_r2) > 0
uspat = uspat[:, :, 0]
data = us_ksp_r2
import pickle
#make the functions for POCS
#=====================================
#number of iterations
numiter = num_iter
# if you want to do an affine data consistency projection
multip = 0 # 0 means simply replacing the measured values
# step size for the prior iterations
alphas = np.logspace(-4, -4, numiter)
# some funtions for simpler coding
def feval(im):
print('im.dtype..............', im.dtype)
return full_funceval(im)
def geval(im):
t1, t2, t3 = full_gradient(im)
return np.reshape(t1, [-1]), np.reshape(t2, [-1]), np.reshape(t3, [-1])
# initialize data with the zero-filled image
recs = np.zeros((imsizer * imrizec, numiter), dtype=complex)
recs[:, 0] = tUFT(data, uspat).flatten().copy()
# if you want to instead continue reconstruction from an existing image
print(' KCT-INFO: contRec is ' + contRec)
if contRec != '':
print('KCT-INFO: reading from a previous file ' + contRec)
rr = pickle.load(open(contRec, 'rb'))
recs[:, 0] = rr[:, -1]
print('KCT-INFO: initialized to the previous recon: ' + contRec)
import time
# the itertaion loop
for it in range(numiter - 1):
start = time.time()
# get the step size for the Piteration
alpha = alphas[it]
# if you want to do some data consistency iterations before starting the prior projections
# this can be helpful when doing recon with multiple coils, e.g. you can do pure SENSE in the beginning...
# or only do phase projections in the beginning.
if it > onlydciter:
# get the gradients for the prior projection and the likelihood values
ftot, f_prior, f_dc = feval(recs[:, it])
gtot, g_prior, g_dc = geval(recs[:, it])
# print("it no: " + str(it) + " f_tot= " + str(ftot) + " f_prior= " + str(-f_prior) + " f_dc (x1e6)= " + str(f_dc/1e6) + " |g_prior|= " + str(np.linalg.norm(g_prior)) + " |g_dc|= " + str(np.linalg.norm(g_dc)) )
print(
"it no: {0}, f_tot= {1:.2f},f_prior= {2:.2f}, f_dc (x1e6)= {3:.2f}, |g_prior|= {4:.2f}, |g_dc|= {5:.2f}"
.format(it, ftot, f_prior, f_dc / 1e6, np.linalg.norm(g_prior),
np.linalg.norm(g_dc)))
# update the image with the prior gradient
recs[:, it + 1] = recs[:, it] - alpha * g_prior
print("............................................")
print(recs[:, it + 1])
# seperate the magnitude from the phase and do the phase projection
tmpa = np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpp = np.angle(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpaf = tmpa.copy().flatten()
if reglmb == 0:
print("KCT-info: skipping phase proj")
tmpptv = tmpp.copy().flatten()
else:
if regtype == 'TV':
tmpptv = tv_proj(tmpp, mu=0.125, lmb=reglmb,
IT=regiter).flatten() #0.1, 15
elif regtype == 'reg2':
tmpptv = reg2_proj(tmpp, alpha=reglmb,
niter=regiter).flatten() #0.1, 15
else:
raise (TypeError)
# combine back the phase and the magnitude
recs[:, it + 1] = tmpaf * np.exp(1j * tmpptv)
else: # the case where you do only data consistency iterations (also iteration 0)
if not it == 0:
print(
'KCT-info: skipping prior proj for the first onlydciters iter.s, doing only phase proj (then maybe DC proj as well) !!!'
)
recs[:, it + 1] = recs[:, it].copy()
# seperate the magnitude from the phase and do the phase projection
tmpa = np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpp = np.angle(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpaf = tmpa.copy().flatten()
if reglmb == 0:
print("KCT-info: skipping phase proj")
tmpptv = tmpp.copy().flatten()
else:
if regtype == 'TV':
tmpptv = tv_proj(tmpp, mu=0.125, lmb=reglmb,
IT=regiter).flatten() #0.1, 15
elif regtype == 'reg2':
tmpptv = reg2_proj(tmpp, alpha=reglmb,
niter=regiter).flatten() #0.1, 15
else:
raise (TypeError)
# combine back the phase and the magnitude
recs[:, it + 1] = tmpaf * np.exp(1j * tmpptv)
print(recs.shape)
#do the DC projection every 'dcprojiter' iterations
if it < onlydciter + 1 or it % dcprojiter == 0: #
tmp1 = UFT(np.reshape(recs[:, it + 1], [imsizer, imrizec]),
(1 - uspat))
tmp2 = UFT(np.reshape(recs[:, it + 1], [imsizer, imrizec]),
(uspat))
tmp3 = data * uspat[:, :, np.newaxis]
#combine the measured data with the projected image affinely (multip=0 for replacing the values)
tmp = tmp1 + multip * tmp2 + (1 - multip) * tmp3
recs[:, it + 1] = tFT(tmp).flatten()
ftot, f_lik, f_dc = feval(recs[:, it + 1])
print(ftot.shape, f_lik.shape, f_dc.shape)
end = time.time()
print('one iteration time is', end - start)
return recs #YourDatasetModuleHere
import re
import subprocess
from Patcher import Patcher
def _get_output(cmd):
proc = subprocess.Popen(cmd,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT)
(stdout, stderr) = proc.communicate()
stdout = stdout.decode().rstrip("\r\n").lstrip("\r\n")
return stdout
os.chdir("..")
patcher = Patcher("README.md", filetype="markdown")
stdout = _get_output(["./x509sak.py"])
text = "\n```\n$ ./x509sak.py\n%s\n```\n" % (stdout)
patcher.patch("summary", text)
commands = []
command_re = re.compile("Begin of cmd-(?P<cmdname>[a-z]+)")
for match in command_re.finditer(patcher.read()):
cmdname = match.groupdict()["cmdname"]
commands.append(cmdname)
for command in commands:
stdout = _get_output(["./x509sak.py", command, "--help"])
text = "\n```\n%s\n```\n" % (stdout)
patcher.patch("cmd-%s" % (command), text)
arg_enum.append(" %s = '%s'," % (enum_name, optchar))
arg_enum.append("")
first = True
for (option, enum_name, requires_parameter) in opts_long:
if first:
arg_enum.append(" %s = 1000," % (enum_name))
first = False
else:
arg_enum.append(" %s," % (enum_name))
arg_enum.append("};")
arg_enum = "\n".join(arg_enum) + "\n"
cmd_def = [" const char *short_options = \"%s\";" % (short_string)]
cmd_def.append(" struct option long_options[] = {")
for (option, enum_name, requires_parameter) in opts_long:
param = "\"%s\"," % (option)
if requires_parameter:
cmd_def.append(" { %-30s required_argument, 0, %s }," %
(param, enum_name))
else:
cmd_def.append(" { %-30s no_argument, 0, %s }," %
(param, enum_name))
cmd_def.append(" { 0 }")
cmd_def.append(" };")
cmd_def = "\n".join(cmd_def) + "\n"
patcher = Patcher("../pgmopts.c")
patcher.patch("help page", help_code)
patcher.patch("command definition enum", arg_enum)
patcher.patch("command definition", cmd_def)
def vaerecon(us_ksp_r2,
uspat,
orim,
i,
result_all,
method,
dcprojiter,
onlydciter=0,
lat_dim=60,
patchsize=28,
contRec='',
lowresmodel=False,
parfact=10,
num_iter=1,
regiter=15,
reglmb=0.05,
regtype='TV'):
def write_Data(result_all):
with open(os.path.join(savepath, "psnr_" + method + ".txt"),
"w+") as f:
#print(len(result_all))
for i in range(len(result_all)):
f.writelines('current image {} PSNR : '.format(i) + str(result_all[i][0]) + \
" SSIM : " + str(result_all[i][1]) + " HFEN : " + str(result_all[i][2]))
f.write('\n')
print('KCT-INFO: contRec is ' + contRec)
print('KCT-INFO: parfact is ' + str(parfact))
#print("............",us_ksp_r2.shape)
#print(sensmaps.shape)
# set parameters
#==============================================================================
np.random.seed(seed=1)
imsizer = us_ksp_r2.shape[0] #252#256#252
imrizec = us_ksp_r2.shape[1] #308#256#308
#print(imsizer,imrizec)
nsampl = 50
#make a network and a patcher to use later
#==============================================================================
x_rec, x_inp, funop, grd0, _, _, _, _, _, _, _, _, _, _, _, _, _, _ = definevae(
lat_dim=lat_dim, patchsize=patchsize, batchsize=parfact * nsampl)
Ptchr = Patcher(imsize=[imsizer, imrizec],
patchsize=patchsize,
step=int(patchsize / 2),
nopartials=True,
contatedges=True)
nopatches = len(Ptchr.genpatchsizes)
print("KCT-INFO: there will be in total " + str(nopatches) + " patches.")
def dconst(us):
#inp: [nx, ny]
#out: [nx, ny]
return np.linalg.norm(np.fft.fft2(us) * uspat - data)**2
def dconst_grad(us):
#inp: [nx, ny]
#out: [nx, ny]
return 2 * np.fft.ifft2(uspat * (np.fft.fft2(us) * uspat - data))
def prior(us):
#inp: [parfact,ps*ps]
#out: parfact
us = np.abs(us)
funeval = funop.eval(feed_dict={x_rec: np.tile(us, (nsampl, 1))}) #
funeval = np.array(np.split(funeval, nsampl,
axis=0)) # [nsampl x parfact x 1]
return np.mean(funeval, axis=0).astype(np.float64)
def prior_grad(us):
#inp: [parfact, ps*ps]
#out: [parfact, ps*ps]
#print('prior_grad',us.dtype,np.max(us.imag))
usc = us.copy()
usabs = np.abs(us)
#print('prior_grad usabs',usabs.dtype,np.max(usabs.imag))
#print('prior_grad usc',usc.dtype,np.max(usc.imag))
#print('usc == us ? :',usc == us)
grd0eval = grd0.eval(feed_dict={x_rec: np.tile(usabs, (nsampl, 1))
}) # ,x_inp: np.tile(usabs,(nsampl,1))
#grd0eval: [500x784]
grd0eval = np.array(np.split(grd0eval, nsampl,
axis=0)) # [nsampl x parfact x 784]
grd0m = np.mean(grd0eval, axis=0) #[parfact,784]
#print('grd0eval.shape,grd0m.shape',grd0eval.shape,grd0m.shape)
#print(grd0m)
grd0m = usc / np.abs(usc) * grd0m
#print('prior_grad grd0m',grd0m.dtype,np.max(grd0m.imag))
return grd0m # .astype(np.float64)
def prior_grad_patches(ptchs):
#inp: [np, ps, ps]
#out: [np, ps, ps]
#takes set of patches as input and returns a set of their grad.s
#both grads are in the positive direction
shape_orig = ptchs.shape
ptchs = np.reshape(ptchs, [ptchs.shape[0], -1])
grds = np.zeros([
int(np.ceil(ptchs.shape[0] / parfact) * parfact),
np.prod(ptchs.shape[1:])
],
dtype=np.complex128)
extraind = int(
np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, ((0, extraind), (0, 0)), mode='edge')
#print('prior_grad_patches',ptchs.dtype,np.max(ptchs.imag))
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
grds[parfact * ix:parfact * ix + parfact, :] = prior_grad(
ptchs[parfact * ix:parfact * ix + parfact, :])
grds = grds[0:shape_orig[0], :]
print('np.reshape(grds, shape_orig).shape',
np.reshape(grds, shape_orig).shape)
return np.reshape(grds, shape_orig)
#np.ceil shang qu zheng
def prior_patches(ptchs):
#inp: [np, ps, ps]
#out: 1
fvls = np.zeros([int(np.ceil(ptchs.shape[0] / parfact) * parfact)])
extraind = int(
np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, [(0, extraind), (0, 0), (0, 0)], mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
fvls[parfact * ix:parfact * ix + parfact] = prior(
np.reshape(ptchs[parfact * ix:parfact * ix + parfact, :, :],
[parfact, -1]))
fvls = fvls[0:ptchs.shape[0]]
return np.mean(fvls)
def full_gradient(image):
#inp: [nx*nx, 1]
#out: [nx, ny], [nx, ny]
#returns both gradients in the respective positive direction.
#i.e. must
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer, imrizec]))
ptchs = np.array(ptchs)
#print('ptchs dtype',ptchs.dtype,np.max(ptchs))
grd_prior = prior_grad_patches(ptchs)
grd_prior = (-1) * Ptchr.patches2im(grd_prior)
grd_dconst = dconst_grad(np.reshape(image, [imsizer, imrizec]))
return grd_prior + grd_dconst, grd_prior, grd_dconst
def full_funceval(image):
#inp: [nx*nx, 1]
#out: [1], [1], [1]
tmpimg = np.reshape(image, [imsizer, imrizec])
dc = dconst(tmpimg)
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer, imrizec]))
ptchs = np.array(ptchs)
priorval = (-1) * prior_patches(np.abs(ptchs))
return priorval + dc, priorval, dc
#define the phase regularization functions
#==============================================================================
def tv_proj(phs, mu=0.125, lmb=2, IT=225):
# Total variation based projection
phs = fb_tv_proj(phs, mu=mu, lmb=lmb, IT=IT)
return phs
def fgrad(im):
# gradient operation with 1st order finite differences
imr_x = np.roll(im, shift=-1, axis=0)
imr_y = np.roll(im, shift=-1, axis=1)
grd_x = imr_x - im
grd_y = imr_y - im
return np.array((grd_x, grd_y))
def fdivg(im):
# divergence operator with 1st order finite differences
imr_x = np.roll(np.squeeze(im[0, :, :]), shift=1, axis=0)
imr_y = np.roll(np.squeeze(im[1, :, :]), shift=1, axis=1)
grd_x = np.squeeze(im[0, :, :]) - imr_x
grd_y = np.squeeze(im[1, :, :]) - imr_y
return grd_x + grd_y
def f_st(u, lmb):
# soft thresholding
uabs = np.squeeze(np.sqrt(np.sum(u * np.conjugate(u), axis=0)))
tmp = 1 - lmb / uabs
tmp[np.abs(tmp) < 0] = 0
uu = u * np.tile(tmp[np.newaxis, :, :], [u.shape[0], 1, 1])
return uu
def fb_tv_proj(im, u0=0, mu=0.125, lmb=1, IT=15):
sz = im.shape
us = np.zeros((2, sz[0], sz[1], IT))
us[:, :, :, 0] = u0
for it in range(IT - 1):
#grad descent step:
tmp1 = im - fdivg(us[:, :, :, it])
tmp2 = mu * fgrad(tmp1)
tmp3 = us[:, :, :, it] - tmp2
#thresholding step:
us[:, :, :, it + 1] = tmp3 - f_st(tmp3, lmb=lmb)
#endfor
return im - fdivg(us[:, :, :, IT - 1])
def reg2_proj(usph, niter=100, alpha=0.05):
#A smoothness based based projection. Regularization method 2 from
#"Separate Magnitude and Phase Regularization via Compressed Sensing", Feng Zhao et al, IEEE TMI, 2012
usph = usph + np.pi * 3 / 2
ims = np.zeros((imsizer, imrizec, niter))
ims[:, :, 0] = usph.copy()
for ix in range(niter - 1):
ims[:, :, ix + 1] = ims[:, :, ix] - 2 * alpha * np.real(
1j * np.exp(-1j * ims[:, :, ix]) *
fdivg(fgrad(np.exp(1j * ims[:, :, ix]))))
return ims[:, :, -1] - np.pi * 3 / 2
#make the data
#===============================
uspat = np.abs(us_ksp_r2) > 0
uspat = uspat[:, :]
data = us_ksp_r2
import pickle
#make the functions for POCS
#=====================================
#number of iterations
numiter = num_iter
# if you want to do an affine data consistency projection
multip = 0 # 0 means simply replacing the measured values
# step size for the prior iterations
alphas = np.logspace(-4, -4, numiter)
#print('step size for the prior iterations',alphas)
#assert False
def feval(im):
return full_funceval(im)
def geval(im):
t1, t2, t3 = full_gradient(im)
return np.reshape(t1, [-1]), np.reshape(t2, [-1]), np.reshape(t3, [-1])
# initialize data with the zero-filled image
recs = np.zeros((imsizer * imrizec, numiter), dtype=complex)
#recs[:,0] = tUFT(data, uspat).flatten().copy()
recs[:, 0] = np.fft.ifft2(data * uspat).flatten().copy()
recs[:, 0] = abs(recs[:, 0])
print(
'Zerosfilled Psnr :',
compare_psnr(255 * np.abs(np.reshape(recs[:, 0], [imsizer, imrizec])),
255 * np.abs(orim),
data_range=255))
#plt.figure(444444444)
#plt.imshow(np.abs(np.reshape(recs[:,0],[imsizer,imrizec])),cmap='gray')
#plt.show()
# if you want to instead continue reconstruction from an existing image
print(' KCT-INFO: contRec is ' + contRec)
if contRec != '':
print('KCT-INFO: reading from a previous file ' + contRec)
rr = pickle.load(open(contRec, 'rb'))
recs[:, 0] = rr[:, -1]
print('KCT-INFO: initialized to the previous recon: ' + contRec)
import time
# the itertaion loop
max_psnr = 0
max_ssim = 0
min_hfen = 100
for it in range(numiter - 1):
start = time.time()
# get the step size for the Piteration
alpha = alphas[it]
# if you want to do some data consistency iterations before starting the prior projections
# this can be helpful when doing recon with multiple coils, e.g. you can do pure SENSE in the beginning...
# or only do phase projections in the beginning.
if it > onlydciter:
# get the gradients for the prior projection and the likelihood values
#ftot, f_prior, f_dc = feval(recs[:,it])
print('recs dtype Input ', recs.dtype)
gtot, g_prior, g_dc = geval(recs[:, it])
# print("it no: " + str(it) + " f_tot= " + str(ftot) + " f_prior= " + str(-f_prior) + " f_dc (x1e6)= " + str(f_dc/1e6) + " |g_prior|= " + str(np.linalg.norm(g_prior)) + " |g_dc|= " + str(np.linalg.norm(g_dc)) )
#print("it no: {0}, f_tot= {1:.2f},f_prior= {2:.2f}, f_dc (x1e6)= {3:.2f}, |g_prior|= {4:.2f}, |g_dc|= {5:.2f}".format(it,ftot,f_prior,f_dc/1e6,np.linalg.norm(g_prior),np.linalg.norm(g_dc)))
# update the image with the prior gradient
recs[:, it + 1] = recs[:, it] - alpha * g_prior
print(
'current Phase prePsnr :',
compare_psnr(
255 *
np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec])),
255 * np.abs(orim),
data_range=255))
print(
'current Phase Real Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).real),
255 * np.abs(orim.real),
data_range=255))
print(
'current Phase Imag Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).imag),
255 * np.abs(orim.imag),
data_range=255))
#plt.figure(1111)
#plt.imshow(np.abs(np.reshape(recs[:,it+1],[imsizer,imrizec])),cmap='gray')
#plt.show()
#print("............................................")
#print(recs[:,it+1])
# seperate the magnitude from the phase and do the phase projection
tmpa = np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpp = np.angle(np.reshape(recs[:, it + 1],
[imsizer, imrizec])) #+(np.pi*3/2)
print(np.max(tmpp), np.min(tmpp))
tmpaf = tmpa.copy().flatten()
#plt.figure(1111)
#plt.imshow(tmpp,cmap='gray')
#plt.show()
if reglmb == 0:
print("KCT-info: skipping phase proj")
tmpptv = tmpp.copy().flatten()
else:
if regtype == 'TV':
tmpptv = tv_proj(tmpp, mu=0.125, lmb=reglmb,
IT=regiter).flatten() #0.1, 15
elif regtype == 'reg2':
tmpptv = reg2_proj(tmpp, alpha=reglmb,
niter=regiter).flatten() #0.1, 15
else:
raise (TypeError)
# combine back the phase and the magnitude
#tmpptv -= (np.pi*3/2)
recs[:, it + 1] = tmpaf * np.exp(1j * tmpptv) #####here
else: # the case where you do only data consistency iterations (also iteration 0)
if not it == 0:
print(
'KCT-info: skipping prior proj for the first onlydciters iter.s, doing only phase proj (then maybe DC proj as well) !!!'
)
recs[:, it + 1] = recs[:, it].copy()
print(
'current Phase prePsnr :',
compare_psnr(
255 *
np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec])),
255 * np.abs(orim),
data_range=255))
print(
'current Phase Real Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).real),
255 * np.abs(orim.real),
data_range=255))
print(
'current Phase Imag Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).imag),
255 * np.abs(orim.imag),
data_range=255))
# seperate the magnitude from the phase and do the phase projection
tmpa = np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpp = np.angle(np.reshape(recs[:, it + 1], [imsizer, imrizec]))
tmpaf = tmpa.copy().flatten()
if reglmb == 0:
print("KCT-info: skipping phase proj")
tmpptv = tmpp.copy().flatten()
else:
if regtype == 'TV':
tmpptv = tv_proj(tmpp, mu=0.125, lmb=reglmb,
IT=regiter).flatten() #0.1, 15
elif regtype == 'reg2':
tmpptv = reg2_proj(tmpp, alpha=reglmb,
niter=regiter).flatten() #0.1, 15
else:
raise (TypeError)
# combine back the phase and the magnitude
recs[:, it + 1] = tmpaf * np.exp(1j * tmpptv)
#plt.figure(22222)
#plt.imshow(np.abs(np.reshape(recs[:,it+1],[imsizer,imrizec])),cmap='gray')
#plt.show()
#do the DC projection every 'dcprojiter' iterations
if it < onlydciter + 1 or it % dcprojiter == 0: #
tmp_rec = np.reshape(recs[:, it + 1], [imsizer, imrizec])
#plt.figure(3333333)
#plt.imshow(np.abs(tmp1),cmap='gray')
#plt.show()
print(np.max(np.abs(tmp_rec)), np.max(255 * np.abs(orim)))
print(
'current Prior Psnr :',
compare_psnr(255 * np.abs(tmp_rec),
255 * np.abs(orim),
data_range=255))
print(
'current Prior Real Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).real),
255 * np.abs(orim.real),
data_range=255))
print(
'current Prior Imag Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).imag),
255 * np.abs(orim.imag),
data_range=255))
#tmp1 = np.fft.fft2(tmp_rec)
#print(uspat)
#tmp1[uspat==1] = us_ksp_r2[uspat==1]
#tmp1 = np.fft.ifft2(tmp1)
#recs[:,it+1] = tmp1.flatten()
#plt.figure(1)
#plt.imshow(np.abs(tmp1),cmap='gray')
#plt.show()
tmp1 = np.fft.fft2(np.reshape(recs[:, it + 1],
[imsizer, imrizec])) * (1 - uspat)
tmp2 = np.fft.fft2(np.reshape(recs[:, it + 1],
[imsizer, imrizec])) * (uspat)
tmp3 = data * uspat #[:,:,np.newaxis]
#combine the measured data with the projected image affinely (multip=0 for replacing the values)
tmp = tmp1 + multip * tmp2 + (1 - multip) * tmp3
recs[:, it + 1] = np.fft.ifft2(tmp).flatten() #np.abs(
print('recs dtype DC ', recs.dtype)
print(
'current DC Psnr :',
compare_psnr(
255 *
np.abs(np.reshape(recs[:, it + 1], [imsizer, imrizec])),
255 * np.abs(orim),
data_range=255))
print(
'current DC Real Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).real),
255 * np.abs(orim.real),
data_range=255))
print(
'current DC Imag Psnr :',
compare_psnr(255 * np.abs(
np.reshape(recs[:, it + 1], [imsizer, imrizec]).imag),
255 * np.abs(orim.imag),
data_range=255))
#ftot, f_lik, f_dc = feval(recs[:,it+1])
#print(ftot.shape,f_lik.shape,f_dc.shape)
x_complex = np.reshape(recs[:, it + 1], [256, 256])
psnr = compare_psnr(255 * abs(x_complex),
255 * abs(orim),
data_range=255)
ssim = compare_ssim(abs(x_complex), abs(orim), data_range=1)
hfen = compare_hfen(abs(x_complex), abs(orim))
print('current %d image %d iteration PSNR:' % (i, it), psnr, ' SSIM :',
ssim, ' HFEN :', hfen)
if max_psnr < psnr:
result_all[i, 0] = psnr
max_psnr = psnr
result_all[31, 0] = sum(result_all[:31, 0]) / 31
savemat(
os.path.join(savepath,
'img_{}_Rec_'.format(i) + method + '.mat'),
{
'data':
np.array(np.reshape(recs[:, it + 1], [256, 256]),
dtype=np.complex)
})
cv2.imwrite(
os.path.join(savepath,
'img_{}_Rec_'.format(i) + method + '.png'),
np.array(255 * np.abs(np.reshape(recs[:, it + 1], [256, 256])),
dtype=np.uint8))
if max_ssim < ssim:
result_all[i, 1] = ssim
max_ssim = ssim
result_all[31, 1] = sum(result_all[:31, 1]) / 31
if min_hfen > hfen:
result_all[i, 2] = hfen
min_hfen = hfen
result_all[31, 2] = sum(result_all[:31, 2]) / 31
write_Data(result_all)
#print('recs dtype DC ',recs.dtype)
end = time.time()
print('{} iteration time is'.format(it), end - start)
return recs #YourDatasetModuleHere
def vaerecon(us_ksp_r2, # undersampled k space
sensmaps,
dcprojiter,
onlydciter = 0,
lat_dim = 60,
patchsize = 28,
contRec = '',
parfact = 10,
num_iter = 302,
rescaled = False,
half = False,
regiter = 15,
reglmb = 0.05,
regtype = 'TV',
usemeth = 1,
stepsize = 1e-4,
optScale = False,
mode = [],
chunks40 = False,
Melmodels = '',
N4BFcorr = False,
z_multip = 1.0,
n1 = 5,
n2 = 5,
log_dir = ''):
logging.info('xxxxxxxxxxxxxxxxxxx contRec is ' + contRec)
logging.info('xxxxxxxxxxxxxxxxxxx parfact is ' + str(parfact) )
# ==============================================================================
# set parameters
# ==============================================================================
np.random.seed(seed = 1)
imsizer = us_ksp_r2.shape[0] #252#256#252
imrizec = us_ksp_r2.shape[1] #308#256#308
nsampl = 50 #0
# ==============================================================================
# make a network and a patcher to use later
# ==============================================================================
# =================================
# get the output from the VAE
# funop: ELBO
# grd0: gradient of the ELBO wrt the image
# grd_p_x_z0, grd_p_z0, grd_q_z_x0: different gradients
# =================================
vae_outputs = definevae(lat_dim = lat_dim,
patchsize = patchsize,
batchsize = parfact*nsampl,
rescaled = rescaled,
half = half,
mode = mode,
chunks40 = chunks40,
Melmodels = Melmodels,
use_normalizer = bool(n1>0),
log_dir = log_dir)
x_rec, x_inp, funop, grd0, sess = vae_outputs[0:5]
grd_p_x_z0, grd_p_z0, grd_q_z_x0, grd20 = vae_outputs[5:9] # these are not being used in the code now.
y_out, y_out_prec, z_std_multip, op_q_z_x = vae_outputs[9:13] # these are not being used in the code now.
mu, std = vae_outputs[13:15] # these are not being used in the code now.
grd_q_zpl_x_az0, op_q_zpl_x, z_pl, z = vae_outputs[15:19] # these are not being used in the code now.
# Most of these outputs were being used in the function likelihood_grad_meth3, which is no longer being used.
norm_accum_grads_zero_op, norm_accum_grads_op, norm_accum_grads_mean_op, num_accum_steps_pl, norm_update_op, x_norm = vae_outputs[19:25]
f_elbo, f_data_consistency, f_total, f_summary = vae_outputs[25:29]
g_elbo, g_data_consistency, g_total, g_summary, summary_writer = vae_outputs[29:34]
# =================================
# used to go from image to patches and back
# =================================
Ptchr = Patcher(imsize = [imsizer, imrizec],
patchsize = patchsize,
step = int(patchsize/2),
nopartials = True,
contatedges = True)
nopatches = len(Ptchr.genpatchsizes)
logging.info("There will be in total " + str(nopatches) + " patches.")
# =================================
# functions for data consistency
# =================================
def dconst(us):
#inp: [nx, ny]
#out: [nx, ny]
return np.linalg.norm(utils.UFT_with_sensmaps(us, uspat, sensmaps) - data)**2
def dconst_grad(us):
#inp: [nx, ny]
#out: [nx, ny]
return 2 * utils.tUFT_with_sensmaps(utils.UFT_with_sensmaps(us, uspat, sensmaps) - data, uspat, sensmaps)
# =================================
# function for running the normalization op on a batch of patches
# =================================
def update_normalizer(im,
sess,
accum_gradients_zero_op,
accum_gradients_op,
accum_gradients_mean_op,
num_accum_steps_pl,
update_normalizer_op):
# make patches
ptchs = Ptchr.im2patches(np.reshape(im, [imsizer, imrizec]))
ptchs = np.array(ptchs)
ptchs = np.abs(ptchs)
# zero accumulated gradients
sess.run(accum_gradients_zero_op)
num_accumulation_steps = 0
# convert image to patches
ptchs = np.reshape(ptchs, [ptchs.shape[0], -1])
# extra indices that have to be padded to ensure we have complete batches for the VAE
extraind = int(np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, ((0, extraind), (0,0)), mode = 'edge')
# send batches of patches and accumulate gradients
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
batch_of_patches = ptchs[parfact * ix : parfact * ix + parfact, :]
sess.run(accum_gradients_op, feed_dict = {x_rec: np.tile(batch_of_patches, (nsampl, 1)), z_std_multip: z_multip})
num_accumulation_steps = num_accumulation_steps + 1
# run op to mean gradients accumulated so far
sess.run(accum_gradients_mean_op, feed_dict = {num_accum_steps_pl: num_accumulation_steps})
# update normalizer parameters according to the mean gradient.
# The stuff passed in the feed_dict does not matter in this line, but something needs to be passed. So passing the last batch of patches.
sess.run(update_normalizer_op, feed_dict = {x_rec: np.tile(batch_of_patches, (nsampl, 1)), z_std_multip: z_multip})
return 0
# =================================
# functions for computing the ELBO and its derivatives
# =================================
# =================================
# returns the elbo of a batch of patches
# =================================
def likelihood(us):
# inp: [parfact,ps*ps]
# out: parfact
us = np.abs(us)
funeval = funop.eval(feed_dict = {x_rec: np.tile(us, (nsampl,1)), z_std_multip: z_multip})
funeval = np.array(np.split(funeval, nsampl, axis=0)) # [nsampl x parfact x 1]
return np.mean(funeval, axis=0).astype(np.float64)
# =================================
# returns the elbo of the input patches
# =================================
def likelihood_patches(ptchs):
# inp: [np, ps, ps]
# out: 1
fvls = np.zeros([int(np.ceil(ptchs.shape[0] / parfact) * parfact) ])
extraind = int(np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, [(0,extraind), (0,0), (0,0)], mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0]/parfact))):
fvls[parfact*ix : parfact*ix+parfact] = likelihood(np.reshape(ptchs[parfact*ix : parfact*ix+parfact,:,:], [parfact,-1]))
fvls = fvls[0:ptchs.shape[0]]
return np.mean(fvls)
# =================================
# returns the ELBO of the image
# =================================
def full_funceval(image):
#inp: [nx*nx, 1]
#out: [1], [1], [1]
tmpimg = np.reshape(image, [imsizer,imrizec])
# how consistent is this image with the measured undersampled k-space data
dc = dconst(tmpimg)
# convert the image into patches and measure the elbo of each patch
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer,imrizec]))
ptchs = np.array(ptchs)
lik = (-1)*likelihood_patches(np.abs(ptchs))
return lik + dc, lik, dc
# =================================
# returns the gradient of the elbo of a batch of patches wrt the batch of patches
# =================================
def likelihood_grad(us):
# inp: [parfact, ps*ps]
# out: [parfact, ps*ps]
usc = us.copy()
usabs = np.abs(us)
grd0eval = grd0.eval(feed_dict = {x_rec: np.tile(usabs, (nsampl, 1)), z_std_multip: z_multip}) # grd0eval: [500x784]
grd0eval = np.array(np.split(grd0eval, nsampl, axis=0)) # [nsampl x parfact x 784]
grd0m = np.mean(grd0eval,axis=0) # [parfact,784]
# correction for the complex image
grd0m = usc/np.abs(usc)*grd0m
return grd0m #.astype(np.float64)
# =================================
# returns the gradient of the elbo of the input patches wrt the input patches
# =================================
def likelihood_grad_patches(ptchs):
# inp: [np, ps, ps]
# out: [np, ps, ps]
# takes set of patches as input and returns a set of their grad.s
# both grads are in the positive direction
shape_orig = ptchs.shape
ptchs = np.reshape(ptchs, [ptchs.shape[0], -1] )
grds = np.zeros([int(np.ceil(ptchs.shape[0]/parfact)*parfact), np.prod(ptchs.shape[1:])], dtype = np.complex64)
# extra indices that have to be padded to ensure we have complete batches for the VAE
extraind = int(np.ceil(ptchs.shape[0] / parfact) * parfact) - ptchs.shape[0]
ptchs = np.pad(ptchs, ((0, extraind), (0,0)), mode='edge')
for ix in range(int(np.ceil(ptchs.shape[0] / parfact))):
grds[parfact*ix : parfact*ix+parfact, :] = likelihood_grad(ptchs[parfact*ix : parfact*ix+parfact, :])
grds = grds[0:shape_orig[0],:]
return np.reshape(grds, shape_orig)
# =================================
# returns the gradient of the ELBO wrt the image
# =================================
def full_gradient(image):
#inp: [nx*nx, 1]
#out: [nx, ny], [nx, ny]
#returns both gradients in the respective positive direction.
#i.e. must
# convert image to patches
ptchs = Ptchr.im2patches(np.reshape(image, [imsizer,imrizec]))
ptchs = np.array(ptchs)
# compute gradient of the ELBO wrt each patch
grd_lik = likelihood_grad_patches(ptchs)
grd_lik = (-1)* Ptchr.patches2im(grd_lik)
grd_dconst = dconst_grad(np.reshape(image, [imsizer,imrizec]))
return grd_lik + grd_dconst, grd_lik, grd_dconst
# =================================
# phase projection functions
# =================================
def tv_proj(phs,mu=0.125,lmb=2,IT=225):
phs = fb_tv_proj(phs,mu=mu,lmb=lmb,IT=IT)
return phs
def fgrad(im):
imr_x = np.roll(im,shift=-1,axis=0)
imr_y = np.roll(im,shift=-1,axis=1)
grd_x = imr_x - im
grd_y = imr_y - im
return np.array((grd_x, grd_y))
def fdivg(im):
imr_x = np.roll(np.squeeze(im[0,:,:]),shift=1,axis=0)
imr_y = np.roll(np.squeeze(im[1,:,:]),shift=1,axis=1)
grd_x = np.squeeze(im[0,:,:]) - imr_x
grd_y = np.squeeze(im[1,:,:]) - imr_y
return grd_x + grd_y
def f_st(u,lmb):
uabs = np.squeeze(np.sqrt(np.sum(u*np.conjugate(u),axis=0)))
tmp = 1 - lmb/uabs
tmp[np.abs(tmp) < 0] = 0
uu = u*np.tile(tmp[np.newaxis,:,:],[u.shape[0],1,1])
return uu
def fb_tv_proj(im, u0=0, mu=0.125, lmb=1, IT=15):
sz = im.shape
us=np.zeros((2,sz[0],sz[1],IT))
us[:,:,:,0] = u0
for it in range(IT-1):
# grad descent step:
tmp1 = im - fdivg(us[:,:,:,it])
tmp2 = mu*fgrad(tmp1)
tmp3 = us[:,:,:,it] - tmp2
# thresholding step:
us[:,:,:,it+1] = tmp3 - f_st(tmp3, lmb=lmb)
return im - fdivg(us[:,:,:,it+1])
def tikh_proj(usph, niter=100, alpha=0.05):
ims = np.zeros((imsizer, imrizec, niter))
ims[:,:,0] = usph.copy()
for ix in range(niter-1):
ims[:,:,ix+1] = ims[:,:,ix] + alpha*2*fdivg(fgrad(ims[:,:,ix]))
return ims[:,:,-1]
def reg2_proj(usph, niter=100, alpha=0.05):
# from Separate Magnitude and Phase Regularization via Compressed Sensing, Feng Zhao
usph = usph + np.pi
ims = np.zeros((imsizer,imrizec,niter))
ims[:,:,0]=usph.copy()
regval = reg2eval(ims[:,:,0].flatten())
for ix in range(niter-1):
ims[:,:,ix+1] = ims[:,:,ix] +alpha*reg2grd(ims[:,:,ix].flatten()).reshape([252,308]) # *alpha*np.real(1j*np.exp(-1j*ims[:,:,ix])* fdivg(fgrad(np.exp( 1j* ims[:,:,ix] ))) )
regval = reg2eval(ims[:,:,ix+1].flatten())
return ims[:,:,-1] - np.pi
def reg2_dcproj(usph, magim, bfestim, niter=100, alpha_reg=0.05, alpha_dc=0.05):
# from Separate Magnitude and Phase Regularization via Compressed Sensing, Feng Zhao
# usph = usph+np.pi
ims = np.zeros((imsizer,imrizec,niter))
grds_reg = np.zeros((imsizer,imrizec,niter))
grds_dc = np.zeros((imsizer,imrizec,niter))
ims[:,:,0]=usph.copy()
regval = reg2eval(ims[:,:,0].flatten())
for ix in range(niter-1):
grd_reg = reg2grd(ims[:,:,ix].flatten()).reshape([252,308]) # *alpha*np.real(1j*np.exp(-1j*ims[:,:,ix])* fdivg(fgrad(np.exp( 1j* ims[:,:,ix] ))) )
grds_reg[:,:,ix] = grd_reg
grd_dc = reg2_dcgrd(ims[:,:,ix].flatten() , magim, bfestim).reshape([252,308])
grds_dc[:,:,ix] = grd_dc
ims[:,:,ix+1] = ims[:,:,ix] + alpha_reg*grd_reg - alpha_dc*grd_dc
regval = reg2eval(ims[:,:,ix+1].flatten())
f_dc = dconst(magim*np.exp(1j*ims[:,:,ix+1])*bfestim)
print_info = False
if print_info is True:
logging.info("norm grad reg: " + str(np.linalg.norm(grd_reg)))
logging.info("norm grad dc: " + str(np.linalg.norm(grd_dc)))
logging.info("regval: " + str(regval))
logging.info("fdc: (*1e9) {0:.6f}".format(f_dc/1e9))
return ims[:,:,-1] #-np.pi
def reg2eval(im):
# takes in 1d, returns scalar
im = im.reshape([252,308])
phs = np.exp(1j*im)
return np.linalg.norm(fgrad(phs).flatten())
def reg2grd(im):
# takes in 1d, returns 1d
im = im.reshape([252,308])
return -2*np.real(1j*np.exp(-1j*im) * fdivg(fgrad(np.exp(1j * im)))).flatten()
def reg2_dcgrd(phim, magim, bfestim):
# takes in 1d, returns 1d
phim = phim.reshape([252,308])
magim = magim.reshape([252,308])
tmp = utils.UFT_with_sensmaps(bfestim * np.exp(1j * phim) * magim, uspat, sensmaps) - data
return -2 * np.real(1j * np.exp(-1j*phim) * magim * bfestim * utils.tUFT_with_sensmaps(tmp, uspat, sensmaps)).flatten()
def reg2_proj_ls(usph, niter=100):
# from Separate Magnitude and Phase Regularization via Compressed Sensing, Feng Zhao
# with line search
usph = usph + np.pi
ims = np.zeros((imsizer, imrizec, niter))
ims[:,:,0] = usph.copy()
regval = reg2eval(ims[:,:,0].flatten())
logging.info(regval)
for ix in range(niter-1):
currgrd = reg2grd(ims[:,:,ix].flatten())
res = sop.minimize_scalar(lambda alpha: reg2eval(ims[:,:,ix].flatten() + alpha * currgrd ), method='Golden')
alphaopt = res.x
logging.info("optimal alpha: " + str(alphaopt) )
ims[:,:,ix+1] = ims[:,:,ix] + alphaopt*currgrd.reshape([252,308])
regval = reg2eval(ims[:,:,ix+1].flatten())
logging.info("regval: " + str(regval))
return ims[:,:,-1]-np.pi
def N4corrf(im):
phasetmp = np.angle(im)
ddimcabs = np.abs(im)
inputImage = sitk.GetImageFromArray(ddimcabs, isVector=False)
corrector = sitk.N4BiasFieldCorrectionImageFilter();
inputImage = sitk.Cast(inputImage, sitk.sitkFloat32)
output = corrector.Execute(inputImage)
N4biasfree_output = sitk.GetArrayFromImage(output)
n4biasfield = ddimcabs/(N4biasfree_output+1e-9)
if np.isreal(im).all():
return n4biasfield, N4biasfree_output
else:
return n4biasfield, N4biasfree_output*np.exp(1j*phasetmp)
# ===============================
# make the data
# ===============================
uspat = np.abs(us_ksp_r2) > 0
uspat = uspat[:,:,0]
data = us_ksp_r2
trpat = np.zeros_like(uspat)
trpat[:, 120:136] = 1
logging.info(uspat)
# =====================================
# initialize counters
# =====================================
numiter = num_iter
multip = 0 # 0.1
alphas = stepsize*np.ones(numiter) # np.logspace(-4,-4,numiter)
# =====================================
# functions for POCS
# =====================================
def feval(im):
return full_funceval(im)
def geval(im):
t1, t2, t3 = full_gradient(im)
return np.reshape(t1,[-1]), np.reshape(t2,[-1]), np.reshape(t3,[-1])
# =====================================
# initialize data
# =====================================
recs = np.zeros((imsizer*imrizec, numiter+30), dtype=complex)
phaseregvals = []
n4biasfields = []
# =====================================
# first image in the 'recs' list is the initial undersampled image
# =====================================
recs[:, 0] = utils.tUFT_with_sensmaps(data, uspat, sensmaps).flatten().copy()
if N4BFcorr:
n4bf, N4bf_image = N4corrf( np.reshape(recs[:,0],[imsizer,imrizec]) )
recs[:,0] = N4bf_image.flatten()
else:
n4bf = 1
# =====================================
# If the optimization is to be continued from a previous run,
# load the final image from the previous optimization as the first image for the current optimization.
# =====================================
logging.info('contRec is ' + contRec)
if contRec != '':
try:
logging.info('Reading from a previous pickle file ' + contRec)
rr = pickle.load(open(contRec, 'rb'))
recs[:, 0] = rr[:, -1]
logging.info('Initialized to the previous recon from pickle: ' + contRec)
except:
logging.info('Reading from a previous numpy file ' + contRec)
rr = np.load(contRec)
recs[:, 0] = rr[:, -1]
logging.info('Initialized to the previous recon from numpy: ' + contRec)
# =====================================
# main loop
# we don't do GD, we do POCS (projection ontol convex sets)
# o. N1 times gradient updates for min_|phi| -ELBO(|x|)
# a. N2 times gradient updates for min_|x| -ELBO(|x|)
# b. do data consistency projection into a set of x such that || Ex - y ||_2^2 = 0.
# c. N gradient updates for the phase image ps: min_px
# =====================================
for it in range(0, numiter-1, 13):
alpha = alphas[it]
# ===============================================
# first do N1 times magnitude prior iterations wrt normalization module
# ===============================================
recstmp = recs[:, it].copy()
for ix in range(n1):
update_normalizer(recstmp,
sess,
norm_accum_grads_zero_op,
norm_accum_grads_op,
norm_accum_grads_mean_op,
num_accum_steps_pl,
norm_update_op)
ftot, f_lik, f_dc = feval(recstmp)
if N4BFcorr:
f_dc = dconst(recstmp.reshape([imsizer, imrizec]) * n4bf)
gtot, g_lik, g_dc = geval(recstmp) # gradient evaluation: total, wrt_vae, wrt_data_consistency
logging.info("----------- updating normalization module, iteration number: " + str(it+ix))
f_summary_msg = sess.run(f_summary, feed_dict = {f_elbo: f_lik, f_data_consistency: f_dc/1e6, f_total: ftot})
summary_writer.add_summary(f_summary_msg, it+ix)
g_summary_msg = sess.run(g_summary, feed_dict = {g_elbo: np.linalg.norm(g_lik), g_data_consistency: np.linalg.norm(g_dc), g_total: np.linalg.norm(g_lik) + np.linalg.norm(g_dc)})
summary_writer.add_summary(g_summary_msg, it+ix)
# recstmp will not change in this loop. Only the normalization module will.
recs[:, it+ix+1] = recstmp.copy()
# ===============================================
# now do N2 times magnitude prior iterations wrt the image itself
# ===============================================
recstmp = recs[:, it + n1].copy()
for ix in range(n2):
ftot, f_lik, f_dc = feval(recstmp)
if N4BFcorr:
f_dc = dconst(recstmp.reshape([imsizer, imrizec]) * n4bf)
gtot, g_lik, g_dc = geval(recstmp) # gradient evaluation : total, wrt_vae, wrt_data_consistency
logging.info("----------- updating image, iteration number: " + str(it+n1+ix))
f_summary_msg = sess.run(f_summary, feed_dict = {f_elbo: f_lik, f_data_consistency: f_dc/1e6, f_total: ftot})
summary_writer.add_summary(f_summary_msg, it+n1+ix)
g_summary_msg = sess.run(g_summary, feed_dict = {g_elbo: np.linalg.norm(g_lik), g_data_consistency: np.linalg.norm(g_dc), g_total: np.linalg.norm(g_lik) + np.linalg.norm(g_dc)})
summary_writer.add_summary(g_summary_msg, it+n1+ix)
# recstmp will change in this loop. The normalization module will stay fixed.
recstmp = recstmp - alpha * g_lik # g_lik
recs[:, it + ix + n1 + 1] = recstmp.copy()
# ===============================================
# Now do a DC projection.... ACTUALLY, skip the DC projection for now.
# ===============================================
logging.info("dummy step, iteration number: " + str(it+n1+n2+1))
recs[:, it + n1 + n2 + 1] = recs[:, it + n1 + n2]
f_summary_msg = sess.run(f_summary, feed_dict = {f_elbo: f_lik, f_data_consistency: f_dc/1e6, f_total: ftot})
summary_writer.add_summary(f_summary_msg, it+n1+n2+1)
g_summary_msg = sess.run(g_summary, feed_dict = {g_elbo: np.linalg.norm(g_lik), g_data_consistency: np.linalg.norm(g_dc), g_total: np.linalg.norm(g_lik) + np.linalg.norm(g_dc)})
summary_writer.add_summary(g_summary_msg, it+n1+n2+1)
# ===============================================
# now do a phase projection
# ===============================================
tmpa = np.abs(np.reshape(recs[:, it + n1 + n2 + 1], [imsizer, imrizec]))
tmpp = np.angle(np.reshape(recs[:, it + n1 + n2 + 1], [imsizer, imrizec]))
tmpatv = tmpa.copy().flatten()
if reglmb == 0:
logging.info("skipping phase proj, iteration number: " + str(it+n1+n2+2))
tmpptv = tmpp.copy().flatten()
else:
logging.info("doing phase proj, iteration number: " + str(it+n1+n2+2))
if regtype == 'TV': # Total variation
tmpptv = tv_proj(tmpp, mu=0.125,lmb=reglmb,IT=regiter).flatten() # 0.1, 15
elif regtype == 'reg2': # Tikhonov
tmpptv = reg2_proj(tmpp, alpha=reglmb, niter=100).flatten() # 0.1, 15
regval = reg2eval(tmpp)
phaseregvals.append(regval)
logging.info("KCT-dbg: phase reg value is " + str(regval))
elif regtype == 'reg2_dc': # Tikhonov, with additional constraint from data consistency
tmpptv = reg2_dcproj(tmpp, tmpa, n4bf, alpha_reg=reglmb, alpha_dc=reglmb, niter=100).flatten()
elif regtype == 'abs':
tmpptv=np.zeros_like(tmpp).flatten()
elif regtype == 'reg2_ls':
tmpptv = reg2_proj_ls(tmpp, niter=regiter).flatten() #0.1, 15
regval = reg2eval(tmpp)
phaseregvals.append(regval)
logging.info("KCT-dbg: phase reg value is " + str(regval))
else:
logging.info("hey mistake!!!!!!!!!!")
# recombine magnitude and updated phase.
recs[:, it + n1 + n2 + 2] = tmpatv*np.exp(1j*tmpptv)
# add summary after phase projection step
recstmp = recs[:, it + n1 + n2 + 2].copy()
ftot, f_lik, f_dc = feval(recstmp)
gtot, g_lik, g_dc = geval(recstmp)
f_summary_msg = sess.run(f_summary, feed_dict = {f_elbo: f_lik, f_data_consistency: f_dc/1e6, f_total: ftot})
summary_writer.add_summary(f_summary_msg, it+n1+n2+2)
g_summary_msg = sess.run(g_summary, feed_dict = {g_elbo: np.linalg.norm(g_lik), g_data_consistency: np.linalg.norm(g_dc), g_total: np.linalg.norm(g_lik) + np.linalg.norm(g_dc)})
summary_writer.add_summary(g_summary_msg, it+n1+n2+2)
# ===============================================
# now do a data consistency projection
# take the measured part of the k space from the measured data and the remaining part of the k space from the updated image.
# ===============================================
logging.info("doing data consistency step, iteration number: " + str(it+n1+n2+3))
if not N4BFcorr:
tmp1 = utils.UFT_with_sensmaps(np.reshape(recs[:, it + n1 + n2 + 2], [imsizer,imrizec]), (1-uspat), sensmaps)
tmp2 = utils.UFT_with_sensmaps(np.reshape(recs[:, it + n1 + n2 + 2], [imsizer,imrizec]), (uspat), sensmaps)
tmp3 = data * uspat[:,:,np.newaxis]
tmp = tmp1 + multip * tmp2 + (1 - multip) * tmp3
recs[:, it + n1 + n2 + 3] = utils.tFT_with_sensmaps(tmp, sensmaps).flatten()
# ftot, f_lik, f_dc = feval(recs[:, it + n1 + n2 + 3])
# logging.info('f_dc (1e6): ' + str(f_dc/1e6) + ' perc: ' + str(100*f_dc/np.linalg.norm(data)**2))
elif N4BFcorr:
n4bf_prev = n4bf.copy()
imgtmp = np.reshape(recs[:, it + n1 + n2 + 2], [imsizer,imrizec]) # biasfree
imgtmp_bf = imgtmp * n4bf_prev # img with bf
n4bf, N4bf_image = N4corrf(imgtmp_bf) # correct the bf, this correction is supposed to be better now.
imgtmp_new = imgtmp * n4bf
n4biasfields.append(n4bf)
tmp1 = utils.UFT_with_sensmaps(imgtmp_new, (1-uspat), sensmaps)
tmp3 = data * uspat[:, :, np.newaxis]
tmp = tmp1 + (1 - multip) * tmp3 # multip=0 by default
recs[:, it + n1 + n2 + 3] = (utils.tFT_with_sensmaps(tmp, sensmaps) / n4bf).flatten()
# ftot, f_lik, f_dc = feval(recs[:, it + n1 + n2 + 3])
# if N4BFcorr: f_dc = dconst(recs[:, it + n1 + n2 + 3].reshape([imsizer,imrizec])*n4bf)
# logging.info('f_dc (1e6): ' + str(f_dc/1e6) + ' perc: ' + str(100*f_dc / np.linalg.norm(data) ** 2))
# add summary after data consistency step
recstmp = recs[:, it + n1 + n2 + 3].copy()
ftot, f_lik, f_dc = feval(recstmp)
gtot, g_lik, g_dc = geval(recstmp)
f_summary_msg = sess.run(f_summary, feed_dict = {f_elbo: f_lik, f_data_consistency: f_dc/1e6, f_total: ftot})
summary_writer.add_summary(f_summary_msg, it+n1+n2+3)
g_summary_msg = sess.run(g_summary, feed_dict = {g_elbo: np.linalg.norm(g_lik), g_data_consistency: np.linalg.norm(g_dc), g_total: np.linalg.norm(g_lik) + np.linalg.norm(g_dc)})
summary_writer.add_summary(g_summary_msg, it+n1+n2+3)
summary_writer.flush()
return recs, 0, phaseregvals, n4biasfields
