def train_D(x_real):
G.train()
D.train()
z = torch.randn(args.batch_size, args.z_dim, 1, 1).to(device)
x_fake = G(z).detach()
# fake image 1d power spectrum
psd1D_img = np.zeros([x_fake.shape[0], N])
for t in range(x_fake.shape[0]):
gen_imgs = x_fake.permute(0, 2, 3, 1)
img_numpy = gen_imgs[t, :, :, :].cpu().detach().numpy()
img_gray = RGB2gray(img_numpy)
fft = np.fft.fft2(img_gray)
fshift = np.fft.fftshift(fft)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
psd1D = (psd1D - np.min(psd1D)) / (np.max(psd1D) - np.min(psd1D))
psd1D_img[t, :] = psd1D
psd1D_img = torch.from_numpy(psd1D_img).float()
psd1D_img = Variable(psd1D_img, requires_grad=True).to(device)
# real image 1d power spectrum
psd1D_rec = np.zeros([x_real.shape[0], N])
for t in range(x_real.shape[0]):
gen_imgs = x_real.permute(0, 2, 3, 1)
img_numpy = gen_imgs[t, :, :, :].cpu().detach().numpy()
img_gray = RGB2gray(img_numpy)
fft = np.fft.fft2(img_gray)
fshift = np.fft.fftshift(fft)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
psd1D = (psd1D - np.min(psd1D)) / (np.max(psd1D) - np.min(psd1D))
psd1D_rec[t, :] = psd1D
psd1D_rec = torch.from_numpy(psd1D_rec).float()
psd1D_rec = Variable(psd1D_rec, requires_grad=True).to(device)
loss_freq = criterion_freq(psd1D_rec, psd1D_img.detach())
x_real_d_logit = D(x_real)
x_fake_d_logit = D(x_fake)
x_real_d_loss, x_fake_d_loss = d_loss_fn(x_real_d_logit, x_fake_d_logit)
gp = gan.gradient_penalty(functools.partial(D),
x_real,
x_fake,
mode=args.gradient_penalty_mode)
D_loss = (x_real_d_loss + x_fake_d_loss
) + gp * args.gradient_penalty_weight + 2 * loss_freq
D.zero_grad()
D_loss.backward()
D_optimizer.step()
return {'d_loss': x_real_d_loss + x_fake_d_loss, 'gp': gp}
def train_G(x_real):
G.train()
D.train()
z = torch.randn(args.batch_size, args.z_dim, 1, 1).to(device)
x_fake = G(z)
x_fake_d_logit = D(x_fake)
g_loss = g_loss_fn(x_fake_d_logit)
# fake image 1d power spectrum
psd1D_img = np.zeros([x_fake.shape[0], N])
for t in range(x_fake.shape[0]):
gen_imgs = x_fake.permute(0, 2, 3, 1)
img_numpy = gen_imgs[t, :, :, :].cpu().detach().numpy()
img_gray = RGB2gray(img_numpy)
fft = np.fft.fft2(img_gray)
fshift = np.fft.fftshift(fft)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
psd1D = (psd1D - np.min(psd1D)) / (np.max(psd1D) - np.min(psd1D))
psd1D_img[t, :] = psd1D
psd1D_img = torch.from_numpy(psd1D_img).float()
psd1D_img = Variable(psd1D_img, requires_grad=True).to(device)
# real image 1d power spectrum
psd1D_rec = np.zeros([x_real.shape[0], N])
for t in range(x_real.shape[0]):
gen_imgs = x_real.permute(0, 2, 3, 1)
img_numpy = gen_imgs[t, :, :, :].cpu().detach().numpy()
img_gray = RGB2gray(img_numpy)
fft = np.fft.fft2(img_gray)
fshift = np.fft.fftshift(fft)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
psd1D = (psd1D - np.min(psd1D)) / (np.max(psd1D) - np.min(psd1D))
psd1D_rec[t, :] = psd1D
psd1D_rec = torch.from_numpy(psd1D_rec).float()
psd1D_rec = Variable(psd1D_rec, requires_grad=True).to(device)
loss_freq = criterion_freq(psd1D_rec, psd1D_img.detach())
datalossBCE.append(loss_freq.data)
G_loss = loss_freq + g_loss
G.zero_grad()
G_loss.backward()
G_optimizer.step()
return {'g_loss': G_loss}
def getpowerspectrum(g):
ghat = np.fft.fft2(g)
ghatshift = fftpack.fftshift(ghat)
# ghatshift = ghat
gabs = np.abs(ghatshift)**2
ps2 = radialProfile.azimuthalAverage(gabs)
return ps2
def FP_profile(im, xcen, ycen, trim_rad=None, mask=0):
"""
calculate the radial profile and include options for using r or r**2
for the profile. might want to display vs. r, but fit voigt profile vs.
r**2. also provide options to trim radius to exclude outside the FOV and
to set a mask value.
Parameters
----------
im : 2D ndarray containing image (usually read in using pyfits)
xcen : float - X position of ring center
ycen : float - Y position of ring center
trim_rad : float - maximum radial extent of profile
mask : int - data value to use for masking bad data
Returns
-------
prof, r : numpy arrays containing radial profile and radii, respectively
"""
prof = azimuthalAverage(im, center=[xcen, ycen], maskval=mask)
r = np.arange(len(prof))
if trim_rad:
prof = prof[0:trim_rad]
r = r[0:trim_rad]
return prof, r
def FP_profile(im, xcen, ycen, trim_rad=None, mask=0):
"""
calculate the radial profile and include options for using r or r**2
for the profile. might want to display vs. r, but fit voigt profile vs.
r**2. also provide options to trim radius to exclude outside the FOV and
to set a mask value.
Parameters
----------
im : 2D ndarray containing image (usually read in using pyfits)
xcen : float - X position of ring center
ycen : float - Y position of ring center
trim_rad : float - maximum radial extent of profile
mask : int - data value to use for masking bad data
Returns
-------
prof, r : numpy arrays containing radial profile and radii, respectively
"""
prof = azimuthalAverage(im, center=[xcen, ycen], maskval=mask)
r = np.arange(len(prof))
if trim_rad:
prof = prof[0:trim_rad]
r = r[0:trim_rad]
return prof, r
def fft1D(imagen):
nimage = (imagen - np.median(imagen)) / np.std(imagen)
nimage = nimage.astype(float)
nsize = 3
image = np.zeros((nimage.shape[0] * nsize, nimage.shape[1] * nsize))
image[0:nimage.shape[0], 0:nimage.shape[1]] = nimage
# image[nimage.shape[0]:,0:nimage.shape[1]] = np.rot90(nimage)
# image[nimage.shape[0]:,nimage.shape[1]:] = np.rot90(np.rot90(nimage))
# image[0:nimage.shape[0],nimage.shape[1]:] = np.rot90(np.rot90(np.rot90(nimage)))
T = float(image.shape[0])
# Take the fourier transform of the image.
F1 = fftpack.fft2(image)
F2 = fftpack.fftshift(F1)
# Calculate a 2D power spectrum
psf2D = np.abs(F2)**2.
# Calculate the azimuthally averaged 1D power spectrum
psf1D = radialProfile.azimuthalAverage(psf2D,
center=(int(T / 2), int(T / 2)))
v = np.arange(len(psf1D)) / T
vmax = psf2D.shape[0] / 2. / T
ii = list(v).index(vmax)
return v[1:ii], psf1D[1:ii]
def FP_profile(im, xcen, ycen, trim_rad=None, mask=0):
"""
wrap calculating the radial profile to include options for using
r or r**2 for the profile. want to display vs. r, but fit voigt
profile vs. r**2. also provide options to trim radius
to exclude outside the FOV and to set a mask value.
"""
prof = azimuthalAverage(im, center=[xcen, ycen], maskval=mask)
r = np.arange(len(prof))
if trim_rad:
prof = prof[0:trim_rad]
r = r[0:trim_rad]
return prof, r
def psd_input_sdens(file_in,nnn,data_type): boxsize = 3.0 boxb = 1.5 ai = np.fromfile(file_in,dtype=data_type) ai = ai.reshape((nnn,nnn))/np.mean(ai)-1.0 ai_crop = ai[nnn/4:nnn*3/4,nnn/4:nnn*3/4] fi = np.fft.fft2(ai_crop) fi = np.fft.fftshift(fi) psdi = np.abs(fi)**2 psdir = radialProfile.azimuthalAverage(psdi) rrr = np.linspace(1,len(psdir),len(psdir))/boxb*2.0*np.pi return rrr,psdir/boxb**2.0/np.max(psdir)
def do_autocorr_power_spectrum(image_for_fft):
'''
This function performs an auto-correlation of an input image and
outputs a normalized 1D azimuthally averaged power as a function of spatial frequency (PSD).
:param image_for_fft: input image for which the user wishes a 1D PSD
:return: a 1D PSD.
'''
Fnoise = fftpack.fft2(image_for_fft) # Fast fourier transform
Fnoise_shifted = fftpack.fftshift(
Fnoise) # Shifting such that m=0 is at the center of the image.
psd2D_noise = np.abs(Fnoise_shifted)**2 # 2D Power map..
psd1D_noise = radialProfile.azimuthalAverage(
psd2D_noise) # azimuthally averaged 1D power profile.
#NOTE_TO_SELF: There is an alternative function for 1D PSD, which i should explore...
psd1d_noise_normed = psd1D_noise / np.max(np.cumsum(
psd1D_noise)) # Normalizing by the cumulative area under 1D PSD.
return psd1d_noise_normed # return normalized 1D PSD.
def update_model(self,data):
### prepare data ###
self.real, sourceD, targetD = self.prepare_image(data)
sourceDC, self.sourceIndex = self.get_domain_code(sourceD)
targetDC, self.targetIndex = self.get_domain_code(targetD)
sourceC, targetC = sourceDC, targetDC
### generate image ###
if self.E is not None:
c_enc, mu, logvar = self.E(self.real,sourceDC)
c_rand = self.sample_latent_code(c_enc.size())
sourceC = torch.cat([sourceDC, c_enc],1)
targetC = torch.cat([targetDC, c_rand],1)
self.fake = self.G(self.real,targetC)
self.cyc = self.G(self.fake,sourceC)
if self.E is not None:
_, mu_enc, _ = self.E(self.fake,targetDC)
if self.opt.lambda_ide > 0:
self.ide = self.G(self.real,sourceC)
### update D ###
self.errDs = []
for i in range(self.opt.d_num):
errD = self.update_D(self.Ds[i], self.Ds_opt[i], self.real.index_select(0,self.sourceIndex[i]), self.fake.index_select(0,self.targetIndex[i]))
self.errDs.append(errD)
### update G ###
self.errGs, self.errKl, self.errCode, errG_total = [], 0, 0, 0
self.G.zero_grad()
#####change:g_loss pure generate loss
g_loss = 0
#######
for i in range(self.opt.d_num):# d_num = of domain number
errG = self.calculate_G(self.Ds[i], self.fake.index_select(0,self.targetIndex[i]))
errG_total += errG
g_loss += errG
self.errGs.append(errG)
self.errCyc = torch.mean(torch.abs(self.cyc-self.real)) * self.opt.lambda_cyc
errG_total += self.errCyc
if self.opt.lambda_ide > 0:
self.errIde = torch.mean(torch.abs(self.ide-self.real)) * self.opt.lambda_ide
errG_total += self.errIde
######change:loss_freq #######
# fake image 1d power spectrum
N = 179
epsilon = 1e-8
psd1D_img = np.zeros([self.fake.shape[0], N])
for t in range(self.fake.shape[0]):
gen_imgs = self.fake.permute(0, 2, 3, 1)
img_numpy = gen_imgs[t, :, :, :].cpu().detach().numpy()
#img_gray = self.RGB2gray(img_numpy)
img_gray = 0.2989 * img_numpy[:, :, 0] + 0.5870 * img_numpy[:, :, 1] + 0.1140 * img_numpy[:, :, 2]
fft = np.fft.fft2(img_gray)
fshift = np.fft.fftshift(fft)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
psd1D = (psd1D - np.min(psd1D)) / (np.max(psd1D) - np.min(psd1D))
psd1D_img[t, :] = psd1D
psd1D_img = torch.from_numpy(psd1D_img).float()
psd1D_img = Variable(psd1D_img, requires_grad=True).to("cuda")
# real image 1d power spectrum
psd1D_rec = np.zeros([self.real.shape[0], N])
for t in range(self.real.shape[0]):
gen_imgs = self.real.permute(0, 2, 3, 1)
img_numpy = gen_imgs[t, :, :, :].cpu().detach().numpy()
#img_gray = self.RGB2gray(img_numpy)
img_gray = 0.2989 * img_numpy[:, :, 0] + 0.5870 * img_numpy[:, :, 1] + 0.1140 * img_numpy[:, :, 2]
fft = np.fft.fft2(img_gray)
fshift = np.fft.fftshift(fft)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
psd1D = (psd1D - np.min(psd1D)) / (np.max(psd1D) - np.min(psd1D))
psd1D_rec[t, :] = psd1D
psd1D_rec = torch.from_numpy(psd1D_rec).float()
psd1D_rec = Variable(psd1D_rec, requires_grad=True).to('cuda')
criterion_freq = nn.BCELoss()
loss_freq = criterion_freq(psd1D_rec, psd1D_img.detach())
loss_freq *= g_loss ####
lambda_freq = 0.5 ###
errG_total += loss_freq * lambda_freq
if self.E is not None:
self.E.zero_grad()
self.errKL = KL_loss(mu,logvar) * self.opt.lambda_kl
errG_total += self.errKL
errG_total.backward(retain_graph=True)
self.G_opt.step()
self.E_opt.step()
self.G.zero_grad()
self.E.zero_grad()
self.errCode = torch.mean(torch.abs(mu_enc - c_rand)) * self.opt.lambda_c
self.errCode.backward()
self.G_opt.step()
else:
errG_total.backward()
self.G_opt.step()
return errD,errG_total ,loss_freq
def faraday_theory_quiet(i_file,j_file,wl_i,wl_j,alpha_file,bands_name,beam=False,polar_mask=False):
print "Computing Cross Correlations for Bands "+str(bands_name)
# radio_file='/data/wmap/faraday_MW_realdata.fits'
# cl_file='/home/matt/wmap/simul_scalCls.fits'
# nside=1024
# npix=hp.nside2npix(nside)
#
# cls=hp.read_cl(cl_file)
# simul_cmb=hp.sphtfunc.synfast(cls,nside,fwhm=0.,new=1,pol=1);
#
# alpha_radio=hp.read_map(radio_file,hdu='maps/phi');
# alpha_radio=hp.ud_grade(alpha_radio,nside_out=nside,order_in='ring',order_out='ring')
# bands=[43.1,94.5]
q_fwhm=[27.3,11.7]
# wl=np.array([299792458./(band*1e9) for band in bands])
# num_wl=len(wl)
# t_array=np.zeros((num_wl,npix))
# q_array=np.zeros((num_wl,npix))
# u_array=np.zeros((num_wl,npix))
# for i in range(num_wl):
# tmp_cmb=rotate_tqu.rotate_tqu(simul_cmb,wl[i],alpha_radio);
# t_array[i],q_array[i],u_array[i]=tmp_cmb
# iqu_band_i=[t_array[0],q_array[0],u_array[0]]
# iqu_band_j=[t_array[1],q_array[1],u_array[1]]
hdu_i=fits.open(i_file)
hdu_j=fits.open(j_file)
alpha_radio=hp.read_map(alpha_file,hdu='maps/phi')
alpha_radio=hp.ud_grade(alpha_radio,1024)
iqu_band_i=hdu_i['stokes iqu'].data
iqu_band_j=hdu_j['stokes iqu'].data
field_pixels=hdu_i['SQUARE PIXELS'].data
hdu_i.close()
hdu_j.close()
iqu_band_i=hp.smoothing(iqu_band_i,pol=1,fwhm=np.sqrt((smoothing_scale)**2-(q_fwhm[0])**2)*np.pi/(180.*60.),lmax=383)
iqu_band_j=hp.smoothing(iqu_band_j,pol=1,fwhm=np.sqrt((smoothing_scale)**2-(q_fwhm[1])**2)*np.pi/(180.*60.),lmax=383)
#alpha_radio=hp.smoothing(alpha_radio,fwhm=np.pi/180.,lmax=383)
const=2.*(wl_i**2-wl_j**2)
Delta_Q=(iqu_band_i[1]-iqu_band_j[1])/const
Delta_U=(iqu_band_i[2]-iqu_band_j[2])/const
alpha_u=alpha_radio*iqu_band_j[2]
alpha_q=-alpha_radio*iqu_band_j[1]
if polar_mask:
P=np.sqrt(iqu_band_j[1]**2+iqu_band_j[2]**2)
bad_pix=np.where( P < .2e-6)
Delta_Q[bad_pix]=0
Delta_U[bad_pix]=0
alpha_u[bad_pix]=0
alpha_q[bad_pix]=0
cross1_array=[]
cross2_array=[]
L=15*(np.pi/180.)
k=np.arange(1,np.round(500*(L/(2*np.pi))))
l=2*np.pi*k/L
Bl_factor=np.repeat(1,len(k))
if beam:
Bl_factor=hp.gauss_beam(np.pi/180.,383)
for field1 in xrange(4):
pix_cmb=field_pixels.field(field1)
nx=np.sqrt(pix_cmb.shape[0])
flat_dq=np.reshape(Delta_Q[pix_cmb],(nx,nx))
flat_du=np.reshape(Delta_U[pix_cmb],(nx,nx))
flat_aq=np.reshape(alpha_q[pix_cmb],(nx,nx))
flat_au=np.reshape(alpha_u[pix_cmb],(nx,nx))
dq_alm=fft.fftshift(fft.fft2(flat_dq,shape=[450,450]))
du_alm=fft.fftshift(fft.fft2(flat_du,shape=[450,450]))
aq_alm=fft.fftshift(fft.fft2(flat_aq,shape=[450,450]))
au_alm=fft.fftshift(fft.fft2(flat_au,shape=[450,450]))
pw2d_qau=np.real(dq_alm*np.conjugate(au_alm))
pw2d_uaq=np.real(du_alm*np.conjugate(aq_alm))
pw1d_qau=radialProfile.azimuthalAverage(pw2d_qau)
pw1d_uaq=radialProfile.azimuthalAverage(pw2d_uaq)
tmp_cl1=pw1d_qau[k.astype(int)-1]*L**2
tmp_cl2=pw1d_uaq[k.astype(int)-1]*L**2
# index=np.where( (np.sqrt(x**2+y**2) <= k[num_k] +1) & ( np.sqrt(x**2 + y**2) >= k[num_k] -1) )
# tmp1= np.sum(pw2d_qau[index])/(np.pi*( (k[num_k]+1)**2 -(k[num_k]-1)**2 ) )
# tmp2= np.sum(pw2d_uaq[index])/(np.pi*( (k[num_k]+1)**2 -(k[num_k]-1)**2 ) )
# tmp_cl1[num_k]=L**2*tmp1
# tmp_cl2[num_k]=L**2*tmp2
cross1_array.append(tmp_cl1/Bl_factor)
cross2_array.append(tmp_cl2/Bl_factor)
cross1=np.mean(cross1_array,axis=0) ##Average over all Cross Spectra
cross2=np.mean(cross2_array,axis=0) ##Average over all Cross Spectra
hp.write_cl('cl_'+bands_name+'_FR_QxaU.fits',cross1)
hp.write_cl('cl_'+bands_name+'_FR_UxaQ.fits',cross2)
return (cross1,cross2)
def faraday_correlate_quiet(i_file,j_file,wl_i,wl_j,alpha_file,bands,beam=False,polar_mask=False):
print "Computing Cross Correlations for Bands "+str(bands)
hdu_i=fits.open(i_file)
hdu_j=fits.open(j_file)
alpha_radio=hp.read_map(alpha_file,hdu='maps/phi')
alpha_radio=hp.ud_grade(alpha_radio,1024)
iqu_band_i=hdu_i['stokes iqu'].data
iqu_band_j=hdu_j['stokes iqu'].data
sigma_i=hdu_i['Q/U UNCERTAINTIES'].data
sigma_j=hdu_j['Q/U UNCERTAINTIES'].data
field_pixels=hdu_i['SQUARE PIXELS'].data
q_fwhm=[27.3,11.7]
noise_const=np.array([36./f for f in q_fwhm])*1e-6
npix=hp.nside2npix(1024)
sigma_i=[noise_const[0]*np.random.normal(0,1,npix),noise_const[1]*np.random.normal(0,1,npix)]
sigma_j=[noise_const[0]*np.random.normal(0,1,npix),noise_const[1]*np.random.normal(0,1,npix)]
iqu_band_i[1]+=sigma_i[0]
iqu_band_i[2]+=sigma_i[1]
iqu_band_j[1]+=sigma_j[0]
iqu_band_j[2]+=sigma_j[1]
hdu_i.close()
hdu_j.close()
iqu_band_i=hp.smoothing(iqu_band_i,pol=1,fwhm=np.sqrt((smoothing_scale)**2-(q_fwhm[0])**2)*np.pi/(180.*60.),lmax=383)
iqu_band_j=hp.smoothing(iqu_band_j,pol=1,fwhm=np.sqrt((smoothing_scale)**2-(q_fwhm[1])**2)*np.pi/(180.*60.),lmax=383)
#alpha_radio=hp.smoothing(alpha_radio,fwhm=np.pi/180.,lmax=383)
const=2.*(wl_i**2-wl_j**2)
Delta_Q=(iqu_band_i[1]-iqu_band_j[1])/const
Delta_U=(iqu_band_i[2]-iqu_band_j[2])/const
alpha_u=alpha_radio*iqu_band_j[2]
alpha_q=-alpha_radio*iqu_band_j[1]
if polar_mask:
P=np.sqrt(iqu_band_j[1]**2+iqu_band_j[2]**2)
bad_pix=np.where( P < .2e-6)
Delta_Q[bad_pix]=0
Delta_U[bad_pix]=0
alpha_u[bad_pix]=0
alpha_q[bad_pix]=0
cross1_array=[]
cross2_array=[]
cross3_array=[]
L=15*(np.pi/180.)
k=np.arange(1,np.round(500*(L/(2*np.pi))))
l=2*np.pi*k/L
Bl_factor=np.repeat(1,len(k))
if beam:
Bl_factor=hp.gauss_beam(np.pi/180.,383)
for field1 in xrange(4):
pix_cmb=field_pixels.field(field1)
nx=np.sqrt(pix_cmb.shape[0])
flat_dq=np.reshape(Delta_Q[pix_cmb],(nx,nx))
flat_du=np.reshape(Delta_U[pix_cmb],(nx,nx))
flat_aq=np.reshape(alpha_q[pix_cmb],(nx,nx))
flat_au=np.reshape(alpha_u[pix_cmb],(nx,nx))
dq_alm=fft.fftshift(fft.fft2(flat_dq,shape=[450,450]))
du_alm=fft.fftshift(fft.fft2(flat_du,shape=[450,450]))
aq_alm=fft.fftshift(fft.fft2(flat_aq,shape=[450,450]))
au_alm=fft.fftshift(fft.fft2(flat_au,shape=[450,450]))
pw2d_qau=np.real(dq_alm*np.conjugate(au_alm))
pw2d_uaq=np.real(du_alm*np.conjugate(aq_alm))
pw1d_qau=radialProfile.azimuthalAverage(pw2d_qau)
pw1d_uaq=radialProfile.azimuthalAverage(pw2d_uaq)
tmp_cl1=pw1d_qau[k.astype(int)-1]*L**2
tmp_cl2=pw1d_uaq[k.astype(int)-1]*L**2
# index=np.where( (np.sqrt(x**2+y**2) <= k[num_k] +1) & ( np.sqrt(x**2 + y**2) >= k[num_k] -1) )
# tmp1= np.sum(pw2d_qau[index])/(np.pi*( (k[num_k]+1)**2 -(k[num_k]-1)**2 ) )
# tmp2= np.sum(pw2d_uaq[index])/(np.pi*( (k[num_k]+1)**2 -(k[num_k]-1)**2 ) )
# tmp_cl1[num_k]=L**2*tmp1
# tmp_cl2[num_k]=L**2*tmp2
cross1_array.append(tmp_cl1/Bl_factor)
cross2_array.append(tmp_cl2/Bl_factor)
cross1=np.mean(cross1_array,axis=0) ##Average over all Cross Spectra
cross2=np.mean(cross2_array,axis=0) ##Average over all Cross Spectra
hp.write_cl('cl_'+bands+'_FR_QxaU.fits',cross1)
hp.write_cl('cl_'+bands+'_FR_UxaQ.fits',cross2)
return (cross1,cross2)
def singleImageAnalysis(model_path):
print("Loading the training set...")
train_img = os.path.join(VanHateren.DATA_DIR, 'train.pkl')
train_set = serial.load(train_img)
patch_size = (32, 32)
# Run the model and visualize
print("Loading the model...")
model = serial.load(model_path)
list_of_images = []
list_of_reconstructed = []
print("Beginning the fft analysis...")
for ii in np.arange(train_set.X.shape[0]):
img_vector = train_set.X[ii, :]
# Part for reconstructed
[tensor_var] = model.reconstruct([img_vector])
reconstructed_vector = tensor_var.eval()
# Save off human-viewable image patches
img_patch = train_set.denormalize_image(img_vector)
img_patch = img_patch.reshape(patch_size)
list_of_images.append(img_patch)
reconstructed_patch = train_set.denormalize_image(reconstructed_vector)
reconstructed_patch = reconstructed_patch.reshape(patch_size)
list_of_reconstructed.append(reconstructed_patch)
average_frequency = fft2AverageOnImageSet(list_of_images)
average_reconstructed = fft2AverageOnImageSet(list_of_reconstructed)
# Show the 2D Power Analysis
fh = plt.figure()
fh.add_subplot(1, 2, 1)
plt.imshow(np.log(average_frequency))
plt.title('Original Images')
fh.add_subplot(1, 2, 2)
plt.imshow(np.log(average_reconstructed))
plt.title('Reconstructed Images')
plt.savefig('fft2D.png')
os.system('eog fft2D.png &')
psd1D = radialProfile.azimuthalAverage(average_frequency)
reconstructed_psd1D = radialProfile.azimuthalAverage(average_reconstructed)
fg = plt.figure()
fg.add_subplot(1, 2, 1)
plt.plot(np.log(psd1D))
plt.title('Original Images')
fg.add_subplot(1, 2, 2)
plt.plot(np.log(reconstructed_psd1D))
plt.title('Reconstructed Images')
plt.savefig('fft1D.png')
os.system('eog fft1D.png &')
A13 = A13 / np.max(A13)
A14 = A14 / np.max(A14)
A15 = A15 / np.max(A15)
A16 = A16 / np.max(A16)
A17 = A17 / np.max(A17)
A18 = A18 / np.max(A18)
A19 = A19 / np.max(A19)
A20 = A20 / np.max(A20)
print A1
sys.exit()
A1F1 = np.fft.fft2(A1)
A1F2 = np.fft.fftshift(A1F1)
A1psd2D = np.abs(A1F2)**2
A1psd1D = radialProfile.azimuthalAverage(A1psd2D)
A2F1 = np.fft.fft2(A2)
A2F2 = np.fft.fftshift(A2F1)
A2psd2D = np.abs(A2F2)**2
A2psd1D = radialProfile.azimuthalAverage(A2psd2D)
A3F1 = np.fft.fft2(A3)
A3F2 = np.fft.fftshift(A3F1)
A3psd2D = np.abs(A3F2)**2
A3psd1D = radialProfile.azimuthalAverage(A3psd2D)
A4F1 = np.fft.fft2(A4)
A4F2 = np.fft.fftshift(A4F1)
A4psd2D = np.abs(A4F2)**2
A4psd1D = radialProfile.azimuthalAverage(A4psd2D)
def compute(d=100,
n=1,
w0=10,
cla="high",
test=False,
output="./",
truedir="./true/",
falsedir="./false/",
number=3000,
feature_num=300):
data = {}
epsilon = 1e-8
N = feature_num
y = []
error = []
if not test:
number_iter = number
psd1D_total = np.zeros([number_iter, N])
label_total = np.zeros([number_iter])
psd1D_org_mean = np.zeros(N)
psd1D_org_std = np.zeros(N)
cont = 0
# fake data
rootdir = falsedir
for subdir, dirs, files in os.walk(rootdir):
for file in files:
filename = os.path.join(subdir, file)
img = cv2.imread(filename, 0)
# windows1 = gauss_window(3, 1.0)
# img = correl2d(img, windows1)
# n = 3
# img = ndimage.maximum_filter(img, (n, n))
# img_sobel = filters.sobel(img)
# img = img + img_sobel
# img = img_sobel
# img_laplace = filters.laplace(img, ksize=3, mask=None)
# img = img + img_laplace
# img = img_laplace
# we crop the center
h = int(img.shape[0] / 3)
w = int(img.shape[1] / 3)
img = img[h:-h, w:-w]
# f = np.fft.fft2(img)
# fshift = np.fft.fftshift(f)
if cla == "bandpass":
fshift = butterworth_bandpass_filter(img, d, w0, n)
else:
fshift = butterworthPassFilter(img, d, n, cla)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
# Calculate the azimuthally averaged 1D power spectrum
points = np.linspace(0, N, num=psd1D.size) # coordinates of a
xi = np.linspace(0, N, num=N) # coordinates for interpolation
interpolated = griddata(points, psd1D, xi, method='cubic')
interpolated /= interpolated[0]
psd1D_total[cont, :] = interpolated
label_total[cont] = 0
cont += 1
if cont == number_iter:
break
if cont == number_iter:
break
for x in range(N):
psd1D_org_mean[x] = np.mean(psd1D_total[:, x])
psd1D_org_std[x] = np.std(psd1D_total[:, x])
## real data
psd1D_total2 = np.zeros([number_iter, N])
label_total2 = np.zeros([number_iter])
psd1D_org_mean2 = np.zeros(N)
psd1D_org_std2 = np.zeros(N)
cont = 0
rootdir2 = truedir
for subdir, dirs, files in os.walk(rootdir2):
for file in files:
filename = os.path.join(subdir, file)
parts = filename.split("/")
img = cv2.imread(filename, 0)
# windows1 = gauss_window(3, 1.0)
# img = correl2d(img, windows1)
# n = 3
# img = ndimage.maximum_filter(img, (n, n))
# img_sobel = filters.sobel(img)
# img = img + img_sobel
# img = img_sobel
# img_laplace = filters.laplace(img, ksize=3, mask=None)
# img = img + img_laplace
# img = img_laplace
# we crop the center
h = int(img.shape[0] / 3)
w = int(img.shape[1] / 3)
img = img[h:-h, w:-w]
# f = np.fft.fft2(img)
# fshift = np.fft.fftshift(f)
if cla == "bandpass":
fshift = butterworth_bandpass_filter(img, d, w0, n)
else:
fshift = butterworthPassFilter(img, d, n, cla)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
# Calculate the azimuthally averaged 1D power spectrum
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
points = np.linspace(0, N, num=psd1D.size) # coordinates of a
xi = np.linspace(0, N, num=N) # coordinates for interpolation
interpolated = griddata(points, psd1D, xi, method='cubic')
interpolated /= interpolated[0]
psd1D_total2[cont, :] = interpolated
label_total2[cont] = 1
cont += 1
if cont == number_iter:
break
if cont == number_iter:
break
for x in range(N):
psd1D_org_mean2[x] = np.mean(psd1D_total2[:, x])
psd1D_org_std2[x] = np.std(psd1D_total2[:, x])
y.append(psd1D_org_mean)
y.append(psd1D_org_mean2)
error.append(psd1D_org_std)
error.append(psd1D_org_std2)
psd1D_total_final = np.concatenate((psd1D_total, psd1D_total2), axis=0)
label_total_final = np.concatenate((label_total, label_total2), axis=0)
data["data"] = psd1D_total_final
data["label"] = label_total_final
output = open(output + 'train_3200_2.pkl', 'wb')
pickle.dump(data, output)
output.close()
print("DATA Saved")
# load feature file
pkl_file = open('train_3200_2.pkl', 'rb')
data = pickle.load(pkl_file)
pkl_file.close()
X = data["data"]
y = data["label"]
num = int(X.shape[0] / 2)
num_feat = X.shape[1]
psd1D_org_0 = np.zeros((num, num_feat))
psd1D_org_1 = np.zeros((num, num_feat))
psd1D_org_0_mean = np.zeros(num_feat)
psd1D_org_0_std = np.zeros(num_feat)
psd1D_org_1_mean = np.zeros(num_feat)
psd1D_org_1_std = np.zeros(num_feat)
cont_0 = 0
cont_1 = 0
# We separate real and fake using the label
for x in range(X.shape[0]):
if y[x] == 0:
psd1D_org_0[cont_0, :] = X[x, :]
cont_0 += 1
elif y[x] == 1:
psd1D_org_1[cont_1, :] = X[x, :]
cont_1 += 1
# We compute statistcis
for x in range(num_feat):
psd1D_org_0_mean[x] = np.mean(psd1D_org_0[:, x])
psd1D_org_0_std[x] = np.std(psd1D_org_0[:, x])
psd1D_org_1_mean[x] = np.mean(psd1D_org_1[:, x])
psd1D_org_1_std[x] = np.std(psd1D_org_1[:, x])
# Plot
x = np.arange(0, num_feat, 1)
fig, ax = plt.subplots(figsize=(10, 6))
ax.plot(x,
psd1D_org_0_mean,
alpha=0.5,
color='red',
label='Fake',
linewidth=2.0)
ax.fill_between(x,
psd1D_org_0_mean - psd1D_org_0_std,
psd1D_org_0_mean + psd1D_org_0_std,
color='red',
alpha=0.2)
ax.plot(x,
psd1D_org_1_mean,
alpha=0.5,
color='blue',
label='Real',
linewidth=2.0)
ax.fill_between(x,
psd1D_org_1_mean - psd1D_org_1_std,
psd1D_org_1_mean + psd1D_org_1_std,
color='blue',
alpha=0.2)
plt.tick_params(axis='x', labelsize=20)
plt.tick_params(axis='y', labelsize=20)
ax.legend(loc='best', prop={'size': 20})
plt.xlabel("Spatial Frequency", fontsize=20)
plt.ylabel("Power Spectrum", fontsize=20)
# plt.show()
plt.savefig(output + 'deepfake_celeb_DF' + cla + '_' + str(d) + '_' +
str(w0) + '_' + str(n) + '.png',
bbox_inches='tight')
num = 3
SVM_r = 0
precision_SVM_r = 0
recall_SVM_r = 0
f1_score_SVM_r = 0
if test:
testset(d, n, w0, cla, output, truedir, falsedir, number, feature_num)
for z in range(num):
try:
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
if test:
pkl_file2 = open(output + 'test_deepfake.pkl', 'rb')
data2 = pickle.load(pkl_file2)
X = data2["data"]
y = data2["label"]
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.4)
else:
pkl_file = open(output + 'train_3200_2.pkl', 'rb')
data = pickle.load(pkl_file)
pkl_file.close()
X = data["data"]
y = data["label"]
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.4)
from sklearn.svm import SVC
svclassifier_r = SVC(C=6.37, kernel='rbf', gamma=0.86)
svclassifier_r.fit(X_train, y_train)
print("train loop " + str(z) + " finished")
# print('Accuracy on test set: {:.3f}'.format(svclassifier_r.score(X_test, y_test)))
y_pred = svclassifier_r.predict(X_test)
n_correct = sum(y_pred == y_test)
# confusion_matrix(y_train,y_pred)
# print("准确率: ", n_correct/len(y_pred))
precision_SVM_r += precision_score(y_test, y_pred)
recall_SVM_r += recall_score(y_test, y_pred)
SVM_r += accuracy_score(y_test, y_pred)
f1_score_SVM_r += f1_score(y_test, y_pred)
except:
num -= 1
print(num)
print("Average SVM_r: " + str(SVM_r / num))
print("Precision SVM_r: " + str(precision_SVM_r / num))
print("Recall SVM_r: " + str(recall_SVM_r / num))
print("F1_score SVM_r: " + str(f1_score_SVM_r / num))
def hemisphericalDifferences(left_model_path, right_model_path, plotting=None):
"""
"""
print("Loading the training set...")
train_img = os.path.join(VanHateren.DATA_DIR, 'train.pkl')
train_set = serial.load(train_img)
patch_size = (32, 32)
# Load both models
print("Loading the models...")
left_model = serial.load(left_model_path)
right_model = serial.load(right_model_path)
# Create empty arrays to hold images
image_patches = {
'orig': [],
left_model: [],
right_model: [],
}
print("Beginning the fft analysis...")
# Go through all the images and add them to the lists
for ii in np.arange(train_set.X.shape[0]):
img_vector = train_set.X[ii, :]
for model in [left_model, right_model]:
# Part for reconstructed
[tensor_var] = model.reconstruct([img_vector])
reconstructed_vector = tensor_var.eval()
reconstructed_patch = train_set.denormalize_image(
reconstructed_vector)
reconstructed_patch = reconstructed_patch.reshape(patch_size)
image_patches[model].append(reconstructed_patch)
# Save off results
img_patch = train_set.denormalize_image(img_vector)
img_patch = img_patch.reshape(patch_size)
image_patches['orig'].append(img_patch)
# Run 2D Analysis
average_frequency = fft2AverageOnImageSet(image_patches['orig'])
average_left_reconstructed = fft2AverageOnImageSet(
image_patches[left_model])
average_right_reconstructed = fft2AverageOnImageSet(
image_patches[right_model])
# Run 1D Analysis
psd1D = radialProfile.azimuthalAverage(average_frequency)
reconstructed_left_psd1D = radialProfile.azimuthalAverage(
average_left_reconstructed)
reconstructed_right_psd1D = radialProfile.azimuthalAverage(
average_right_reconstructed)
# Get difference of differences
left_difference = abs(reconstructed_left_psd1D - psd1D)
right_difference = abs(reconstructed_right_psd1D - psd1D)
total_difference = left_difference - right_difference # RH better: > 0
# Plot 3 1D images: The Original, the reconstructions from LH and RH
if plotting:
fig = plt.figure(figsize=(12, 6))
fig.add_subplot(1, 2, 1)
plt.plot(
np.asarray([
np.log(psd1D),
np.log(reconstructed_left_psd1D),
np.log(reconstructed_right_psd1D),
]).T)
plt.legend([
'Original',
'LH (\sigma=%.2f)' % np.asarray(left_model.sigma).max(),
'RH (\sigma=%.2f)' % np.asarray(right_model.sigma).max()
])
plt.xlabel('Spatial frequency')
plt.ylabel('Power (log(amplitude))')
fig.add_subplot(1, 2, 2)
plt.plot(total_difference)
plt.hold(True)
plt.plot(np.ndarray(len(total_difference)),
np.zeros(total_difference.shape)) # show X-axis
plt.title('Closeness in power differences (RH - LH)')
plt.xlabel('Spatial frequency')
plt.ylabel('Power difference')
if isinstance(plotting, basestring):
plt.savefig(plotting)
else:
plt.show()
return total_difference
lam = '70'
num, xx, yy = np.loadtxt('target.txt', unpack=True)
#images = glob.glob('*.fits')
i = 0
f = open('photauto.txt', 'w')
for i in np.arange(len(num)):
if num[i] < 100:
str_num = '0' + str(int(num[i]))
else:
str_num = str(int(num[i]))
data, hdr = fits.getdata('destripe_l' + str_num + '_blue_wgls_rcal.fits',
0,
header=True)
delt = abs(hdr['CDELT1'])
a = radialProfile.azimuthalAverage(data, center=[int(xx[i]), int(yy[i])])
grid = annulus_index.go([data.shape[1], data.shape[0]],
[int(xx[i]), int(yy[i])], 25, 10)
bg = np.median(data[grid])
if np.size(np.where(a - bg < 0.1 * (np.max(a[0:5]) - bg))) == 0:
print >> f, str_num, 'none'
continue
radius = np.min([
np.where(a - bg < 0.1 * (np.max(a[0:5]) - bg))[0][0],
np.where(a <= 2 * bg)[0][0]
])
print radius, radius * abs(hdr['CDELT1']) * 3600.
position = (int(xx[i]), int(yy[i]))
aperture = CircularAperture(position, r=radius)
f = np.fft.fft2(img)
fshift = np.fft.fftshift(f)
print("FFT", f)
print("FFT shift", fshift)
# print(len(fshift), len(fshift[0]))
filtr = [0] * 100
#
# print(2000*np.log(np.abs(fshift)))
# print(filtr)
# fshift = fshift * filtr
magnitude_spectrum2 = np.abs(fshift)**2
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = rp.azimuthalAverage(magnitude_spectrum)
print(psd1D)
py.figure(1)
py.clf()
py.imshow(img)
py.figure(2)
py.clf()
py.imshow(magnitude_spectrum)
py.figure(3)
py.clf()
py.semilogy(psd1D)
py.xlabel("Spatial Frequency")
py.ylabel("Power Spectrum")
py.show()
def getpowerspectrum(g):
gback = fftpack.fft2(g)
gback2 = fftpack.fftshift(gback)
gback2 = np.abs(gback2)**2
ps2 = radialProfile.azimuthalAverage(gback2)
return ps2
def pwspin(NRO, Nup, a, mu0, rgrid, Lpow, deriv, seqnum, data, offsetx, offsety, vmx):
### Universal constants read from file
###number of geodesics
nn=len(data)/(Nup+1)
###total length of padded image
Npad=NRO*10.0#int((nn)/20)
###start of nonzero entries in padded image
cor=int(Npad/2) - int(NRO/2)
###outer edge of accretion disk
rout=15.0#+seqnum/float(seqlen)*12.0
### file specific constants
ufa=zeros(nn)
alpha=zeros(nn)
beta=zeros(nn)
### Luminosity as a function of Impact Params
Lra=zeros(( (Npad),(Npad) ))
##########################
#Fill final radius arrays
for i in range (int((nn-1))):
ufa[i] = data[Nup+i*(Nup+1)][0]
#Fill alpha and beta arrays
for i in range (int((nn-1))):
alpha[i] = data[i*(Nup+1)][0]
beta[i] = data[i*(Nup+1)][1]
##########################
#Define Kerr parameters for ISCO (prograde only!)
r = 1.0/ufa
M = 1.0
###Define margianlly stable orbits in Kerr spacetime#######
Z1 = 1+(1-a**2/M**2)**(1.E0/3.E0) * ( ( 1+a/M )**(1.E0/3.E0) + ( 1-a/M )**(1.E0/3.E0) )
Z2 = ( 3*(a/M)**2+Z1**2 )**(1.E0/2.E0)
rms = M*( 3+Z2 - ( (3-Z1)*(3+Z1+2*Z2) )**(1.E0/2.E0) )
C3 = r>=rms
####RELATIVISTIC BEAMING PARAMS#############
spec_indx = 0.7 #synchrotron spectral index
phi = arctan(alpha/beta) #angle around top of brightness profile
for i in range (0,len(alpha)): #fix arctan symmetries
if (beta[i]>=0.0):
phi[i]=-phi[i]
###vmx is a user defind parameter = v/c which is passed to PWleastsqrs.py
vmax = vmx*sin(acos(mu0)) #Fraction of speed of light of disk rotation along LOS
vr = (rms/r)**(1.0/2.0)*vmax #keplerian Disk has v prop to 1/sqrt(r)
Bet = vr * numpy.sin( phi ) #Beta=v/c as a function of azimuth (phi)
###Define L(r) with REL BEAMING ##########################
for i in range( int(cor), int(cor + NRO) ):
for j in range( int(cor), int(cor + NRO) ):
if (r[(j-cor)+(i-cor)*NRO]<=rout and C3[(j-cor)+(i-cor)*NRO]):
#BEAMING EQUATION: L_{obs}=L_{instrinsic}*sqrt[(1+B)/(1-B)]^(3-spec_indx)
Lra[j][i] = ( (ufa[(j-cor)+(i-cor)*NRO])**( Lpow ) ) * (sqrt( (1+Bet[(j-cor)+(i-cor)*NRO])/(1-Bet[(j-cor)+(i-cor)*NRO]) ))**(3.0-spec_indx)
###Define L(r) without REL BEAMING ##########################
# for i in range( int(cor), int(cor + NRO) ):
# for j in range( int(cor), int(cor + NRO) ):
# # Now asign intrinsic luminosity for photons which eminated
# # from region between rms and rout
# if (r[(j-cor)+(i-cor)*NRO]<=rout and C3[(j-cor)+(i-cor)*NRO]):
# Lra[j+offsety][i+offsetx] = (ufa[(j-cor)+(i-cor)*NRO])**( Lpow )
###############################################################
#################Some previous Luminosty profiles #####################
#3.0*M*Md/(2.0*(r[(j-cor)+(i-cor)*NRO])**3)*( 1-B*(rms/r[(j-cor)+(i-cor)*NRO])**(1.E0/2.E0) )
#(ufa[(j-cor)+(i-cor)*NRO])**( Lpow )
### NY 2008 ADAF two temp ###
## for i in range( int(cor), int(cor + NRO) ):
## for j in range( int(cor), int(cor + NRO) ):
## if (r[(j-cor)+(i-cor)*NRO]<=rout and C3[(j-cor)+(i-cor)*NRO]):
## if (10.0*ufa[(j-cor)+(i-cor)*NRO] > 1.0):
## Lra[j][i] = 10.0*(ufa[(j-cor)+(i-cor)*NRO])**( Lpow ) + 1.0
## if (10.0*ufa[(j-cor)+(i-cor)*NRO] <= 1.0):
## Lra[j+offset][i+offset] = 20.0*(ufa[(j-cor)+(i-cor)*NRO])**( Lpow )
#####################################################################3
#########################
### FFT ###
IfLa = fft2(Lra)
########################
### shift FFT to be centerad at 0 ###
IfLas=fftpack.fftshift(IfLa)
### Get rid of extra zeros in 2-d ifft ###
ii = int(-1)
jj = int(0)
IfLas_n=zeros( ((NRO),(NRO)), 'complex')
for i in range( int(cor), int(cor + NRO) ):
ii = ii+1
for j in range( int(cor), int(cor + NRO) ):
IfLas_n[jj][ii] = IfLas[j][i]
jj = jj+1
if (j==int(cor + NRO-1)):
jj=0
### imaginary part ##
IfaIm = IfLas_n.imag
#########################
### compute power (sqrd ifft) ###
pwIm = abs(IfaIm)**2
If2mag = abs(IfLas_n)**2
########################
### Azimuthal Average ###
pw1dIm = radialProfile.azimuthalAverage(pwIm)
pw1dmag = radialProfile.azimuthalAverage(If2mag)
###########################
### ANGULAR AVG ###
pwang=radialProfile.angularAverage(pwIm)
## SMOOTH ###
from savitzky_golay import *
numpnts = 3
polydeg = 2
coeff = calc_coeff(numpnts, polydeg, diff_order=deriv)
pwIm_smooth = smooth(pw1dIm, coeff)
### NORMALIZE POW ###
pwImmx = max(pwIm_smooth)
pwmagmx = max(pw1dmag)
pwIm_smooth = pwIm_smooth/pwmagmx
pwang = pwang/pwmagmx
pwIm = pwIm/pwmagmx
###########################
### The smoothed azimuthally averaged imaginary power spectrum is returned ###
return pwIm_smooth
def hemisphericalDifferences(left_model_path, right_model_path, plotting=None):
"""
"""
print("Loading the training set...")
train_img = os.path.join(VanHateren.DATA_DIR, 'train.pkl')
train_set = serial.load(train_img)
patch_size = (32, 32)
# Load both models
print("Loading the models...")
left_model = serial.load(left_model_path)
right_model = serial.load(right_model_path)
# Create empty arrays to hold images
image_patches = {
'orig': [],
left_model: [],
right_model: [], }
print("Beginning the fft analysis...")
# Go through all the images and add them to the lists
for ii in np.arange(train_set.X.shape[0]):
img_vector = train_set.X[ii, :]
for model in [left_model, right_model]:
# Part for reconstructed
[tensor_var] = model.reconstruct([img_vector])
reconstructed_vector = tensor_var.eval()
reconstructed_patch = train_set.denormalize_image(reconstructed_vector)
reconstructed_patch = reconstructed_patch.reshape(patch_size)
image_patches[model].append(reconstructed_patch)
# Save off results
img_patch = train_set.denormalize_image(img_vector)
img_patch = img_patch.reshape(patch_size)
image_patches['orig'].append(img_patch)
# Run 2D Analysis
average_frequency = fft2AverageOnImageSet(image_patches['orig'])
average_left_reconstructed = fft2AverageOnImageSet(
image_patches[left_model])
average_right_reconstructed = fft2AverageOnImageSet(
image_patches[right_model])
# Run 1D Analysis
psd1D = radialProfile.azimuthalAverage(average_frequency)
reconstructed_left_psd1D = radialProfile.azimuthalAverage(
average_left_reconstructed)
reconstructed_right_psd1D = radialProfile.azimuthalAverage(
average_right_reconstructed)
# Get difference of differences
left_difference = abs(reconstructed_left_psd1D - psd1D)
right_difference = abs(reconstructed_right_psd1D - psd1D)
total_difference = left_difference - right_difference # RH better: > 0
# Plot 3 1D images: The Original, the reconstructions from LH and RH
if plotting:
fig = plt.figure(figsize=(12, 6))
fig.add_subplot(1, 2, 1)
plt.plot(np.asarray([np.log(psd1D),
np.log(reconstructed_left_psd1D),
np.log(reconstructed_right_psd1D), ]).T)
plt.legend(['Original',
'LH (\sigma=%.2f)' % np.asarray(left_model.sigma).max(),
'RH (\sigma=%.2f)' % np.asarray(right_model.sigma).max()])
plt.xlabel('Spatial frequency')
plt.ylabel('Power (log(amplitude))')
fig.add_subplot(1, 2, 2)
plt.plot(total_difference)
plt.hold(True)
plt.plot(np.ndarray(len(total_difference)),
np.zeros(total_difference.shape)) # show X-axis
plt.title('Closeness in power differences (RH - LH)')
plt.xlabel('Spatial frequency')
plt.ylabel('Power difference')
if isinstance(plotting, basestring):
plt.savefig(plotting)
else:
plt.show()
return total_difference
def singleImageAnalysis(model_path):
print("Loading the training set...")
train_img = os.path.join(VanHateren.DATA_DIR, 'train.pkl')
train_set = serial.load(train_img)
patch_size = (32, 32)
# Run the model and visualize
print("Loading the model...")
model = serial.load(model_path)
list_of_images = []
list_of_reconstructed = []
print("Beginning the fft analysis...")
for ii in np.arange(train_set.X.shape[0]):
img_vector = train_set.X[ii, :]
# Part for reconstructed
[tensor_var] = model.reconstruct([img_vector])
reconstructed_vector = tensor_var.eval()
# Save off human-viewable image patches
img_patch = train_set.denormalize_image(img_vector)
img_patch = img_patch.reshape(patch_size)
list_of_images.append(img_patch)
reconstructed_patch = train_set.denormalize_image(reconstructed_vector)
reconstructed_patch = reconstructed_patch.reshape(patch_size)
list_of_reconstructed.append(reconstructed_patch)
average_frequency = fft2AverageOnImageSet(list_of_images)
average_reconstructed = fft2AverageOnImageSet(list_of_reconstructed)
# Show the 2D Power Analysis
fh = plt.figure()
fh.add_subplot(1, 2, 1)
plt.imshow(np.log(average_frequency))
plt.title('Original Images')
fh.add_subplot(1, 2, 2)
plt.imshow(np.log(average_reconstructed))
plt.title('Reconstructed Images')
plt.savefig('fft2D.png')
os.system('eog fft2D.png &')
psd1D = radialProfile.azimuthalAverage(average_frequency)
reconstructed_psd1D = radialProfile.azimuthalAverage(average_reconstructed)
fg = plt.figure()
fg.add_subplot(1, 2, 1)
plt.plot(np.log(psd1D))
plt.title('Original Images')
fg.add_subplot(1, 2, 2)
plt.plot(np.log(reconstructed_psd1D))
plt.title('Reconstructed Images')
plt.savefig('fft1D.png')
os.system('eog fft1D.png &')
else:
count = count + 1
fig = plt.figure()
ax = fig.add_subplot(111)
ax.imshow(im)
plt.show()
print "Num of Vertical = ", vcount
print "Num of Horizontal = ", hcount
print "Num of Neither = ", fcount
F1 = np.fft.fft2(im)
F2 = np.fft.fftshift(F1)
psd2D = np.abs(F2)**2
psd1D = radialProfile.azimuthalAverage(psd2D)
fig = plt.figure()
ax = fig.add_subplot(111)
ax.imshow(np.log10(im), cmap=py.cm.Greys)
ax.set_ylabel('Vertical=%d\n Horizontal=%d\nNeither=%d' %
(vcount, hcount, fcount),
rotation='horizontal')
fig2 = plt.figure()
ax2 = fig2.add_subplot(111)
ax2.imshow(np.log10(psd2D))
fig3 = plt.figure()
ax3 = fig3.add_subplot(111)
ax3.semilogy(psd1D)
continue
img = cv2.imread(filename,0)
# we crop the center
#h = int(img.shape[0]/3)
#w = int(img.shape[1]/3)
#img = img[h:-h,w:-w]
f = np.fft.fft2(img)
fshift = np.fft.fftshift(f)
magnitude_spectrum = 20*np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
# Calculate the azimuthally averaged 1D power spectrum
points = np.linspace(0,N,num=psd1D.size) # coordinates of a
xi = np.linspace(0,N,num=N) # coordinates for interpolation
interpolated = griddata(points,psd1D,xi,method='cubic')
#cv2.imshow("image", interpolated)
#cv2.waitKey(0)
interpolated /= interpolated[0]
psd1D_total[cont,:] = interpolated
label_total[cont] = 0
def main():
if (len(sys.argv) != 5):
print("Not enough arguments")
print("insert <dir> <features> <max_files> <output filename>")
exit()
dir = sys.argv[1]
N = int(sys.argv[2])
number_iter = int(sys.argv[3])
output_filename = str(sys.argv[4]) + ".pkl"
if os.path.isdir(dir) is False:
print("this directory does not exist")
exit(0)
data = {}
psd1D_total = np.zeros([number_iter, N])
label_total = np.zeros([number_iter])
psd1D_org_mean = np.zeros(N)
psd1D_org_std = np.zeros(N)
cont = 0
rootdir = dir
for subdir, dirs, files in os.walk(rootdir):
#print(files)
for file in files:
#print("entrey boy fake")
filename = os.path.join(subdir, file)
if filename == dir + "\desktop.ini":
continue
img = cv2.imread(filename, 0)
f = np.fft.fft2(img)
fshift = np.fft.fftshift(f)
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
# Calculate the azimuthally averaged 1D power spectrum
points = np.linspace(0, N, num=psd1D.size) # coordinates of a
xi = np.linspace(0, N, num=N) # coordinates for interpolation
interpolated = griddata(points, psd1D, xi, method='cubic')
interpolated /= interpolated[0]
psd1D_total[cont, :] = interpolated
label_total[cont] = 0 #fake
cont += 1
if cont == number_iter:
break
if cont == number_iter:
break
for x in range(N):
psd1D_org_mean[x] = np.mean(psd1D_total[:, x])
psd1D_org_std[x] = np.std(psd1D_total[:, x])
data["data"] = psd1D_total
data["label"] = label_total
output = open(output_filename, 'wb')
pickle.dump(data, output)
output.close()
print("DATA Saved")
def testset(d=100,
n=1,
w0=10,
cla="high",
output="./",
truedir="./true/",
falsedir="./false/",
number=3000,
feature_num=300):
data = {}
epsilon = 1e-8
N = feature_num
y = []
error = []
number_iter = number
psd1D_total = np.zeros([number_iter, N])
label_total = np.zeros([number_iter])
psd1D_org_mean = np.zeros(N)
psd1D_org_std = np.zeros(N)
cont = 0
# fake data
rootdir = falsedir
for subdir, dirs, files in os.walk(rootdir):
for file in files:
filename = os.path.join(subdir, file)
img = cv2.imread(filename, 0)
# we crop the center
h = int(img.shape[0] / 3)
w = int(img.shape[1] / 3)
img = img[h:-h, w:-w]
f = np.fft.fft2(img)
fshift = np.fft.fftshift(f)
if cla == "bandpass":
fshift = butterworth_bandpass_filter(img, d, w0, n)
else:
fshift = butterworthPassFilter(img, d, n, cla)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
# Calculate the azimuthally averaged 1D power spectrum
points = np.linspace(0, N, num=psd1D.size) # coordinates of a
xi = np.linspace(0, N, num=N) # coordinates for interpolation
interpolated = griddata(points, psd1D, xi, method='cubic')
interpolated /= interpolated[0]
psd1D_total[cont, :] = interpolated
label_total[cont] = 0
cont += 1
if cont == number_iter:
break
if cont == number_iter:
break
for x in range(N):
psd1D_org_mean[x] = np.mean(psd1D_total[:, x])
psd1D_org_std[x] = np.std(psd1D_total[:, x])
## real data
psd1D_total2 = np.zeros([number_iter, N])
label_total2 = np.zeros([number_iter])
psd1D_org_mean2 = np.zeros(N)
psd1D_org_std2 = np.zeros(N)
cont = 0
rootdir2 = truedir
for subdir, dirs, files in os.walk(rootdir2):
for file in files:
filename = os.path.join(subdir, file)
parts = filename.split("/")
img = cv2.imread(filename, 0)
# we crop the center
h = int(img.shape[0] / 3)
w = int(img.shape[1] / 3)
img = img[h:-h, w:-w]
f = np.fft.fft2(img)
fshift = np.fft.fftshift(f)
if cla == "bandpass":
fshift = butterworth_bandpass_filter(img, d, w0, n)
else:
fshift = butterworthPassFilter(img, d, n, cla)
fshift += epsilon
magnitude_spectrum = 20 * np.log(np.abs(fshift))
# Calculate the azimuthally averaged 1D power spectrum
psd1D = radialProfile.azimuthalAverage(magnitude_spectrum)
points = np.linspace(0, N, num=psd1D.size) # coordinates of a
xi = np.linspace(0, N, num=N) # coordinates for interpolation
interpolated = griddata(points, psd1D, xi, method='cubic')
interpolated /= interpolated[0]
psd1D_total2[cont, :] = interpolated
label_total2[cont] = 1
cont += 1
if cont == number_iter:
break
if cont == number_iter:
break
for x in range(N):
psd1D_org_mean2[x] = np.mean(psd1D_total2[:, x])
psd1D_org_std2[x] = np.std(psd1D_total2[:, x])
y.append(psd1D_org_mean)
y.append(psd1D_org_mean2)
error.append(psd1D_org_std)
error.append(psd1D_org_std2)
psd1D_total_final = np.concatenate((psd1D_total, psd1D_total2), axis=0)
label_total_final = np.concatenate((label_total, label_total2), axis=0)
data["data"] = psd1D_total_final
data["label"] = label_total_final
output = open(output + 'test_deepfake.pkl', 'wb')
pickle.dump(data, output)
output.close()
print("Deepfake DATA Saved")
#print "13", count13
#print "14", count14
#print "15", count15
#print "16", count16
print AW
F1 = np.fft.fft2(AW)
F2 = np.fft.fftshift(F1)
psd2D = np.abs(F2)**2
psd1D = radialProfile.azimuthalAverage(psd2D)
fig = plt.figure()
ax = fig.add_subplot(111)
ax.imshow(np.log10(AW), cmap=py.cm.Greys)
fig2 = plt.figure()
ax2 = fig2.add_subplot(111)
ax2.imshow(np.log10(psd2D))
fig3 = plt.figure()
ax3 = fig3.add_subplot(111)
ax3.semilogy(psd1D)
ax3.set_xlabel('Spatial Frequency')
F2Lm = fftpack.fftshift(fftLm)
F2S = fftpack.fftshift(fftS)
#Not sure why this is squared www.astrobetter.com/fourier-transforms-of-images-in-python/
#Or why a log is take before plotting, from the same source
F3Lum = abs(F2Lum)**2
F3Lm = abs(F2Lm)**2
F3S = abs(F2S)**2
#Plotting kspace
# pylab.figure(1)
# pylab.clf()
# pylab.imshow(np.log10(F3Lum))
#Calculating the azimuthally average 1D power spectrum using radialProfile
A1Lum = radialProfile.azimuthalAverage(F3Lum)
A1Lm = radialProfile.azimuthalAverage(F3Lm)
A1S = radialProfile.azimuthalAverage(F3S)
#Plot radial Profile
# pylab.figure(2)
# pylab.clf()
# pylab.semilogy(A1Lum, label = 'Lum')
# pylab.semilogy(A1Lm, label = 'Lm')
# pylab.semilogy(A1S, label = 'S')
# pylab.title(OrigImage)
# pylab.legend()
# pylab.xlim(0, 256)
# pylab.xlabel('Spatial Frequency')
# pylab.ylabel('Power Spectrum')
#
