def prepare_data(self,idxs=None):
if self.dataset is 'kth':
print 'loading from KTH dataset'
x_train0, x_test0, y_train0, y_test0 = \
get_kth_imgs(N=50000,n=self.n,reCalc=False,resize=self.original_dim)
x_train0[0],_ = synth2(224,H=0.3)
x_train0[1],_ = synth2(224,H=0.5)
for i in range(2):
x_train0[i]-=np.min(x_train0[i]*1.0)
x_train0[i]=x_train0[i]/np.max(x_train0[i])*256.0
print 'two first samples are fBm'
elif self.dataset is 'other': # fbm
print 'loading from OTHER (fbm) dataset'
x_train0, x_test0, y_train0, y_test0 = \
get_other_imgs(N=1000,n=self.n,reCalc=False,resize=self.original_dim)
#print 'NOT IMPLEMENTED'
elif self.dataset is 'kurtsim':
print 'loading from kurt (sim) dataset'
x_train0, x_test0, y_train0, y_test0 = \
get_kurtsim_imgs(N=120,n=self.n,reCalc=False,resize=self.original_dim)
else:
print 'NOT IMPLEMENTED'
if self.dataset is 'kth' and idxs is not None:
return x_train0[idxs], x_test0, y_train0[idxs], y_test0
else:
return x_train0, x_test0, y_train0, y_test0
def get_G_kth():
texture = Texture()
original_dim = 224**2
n=200
x_train0, x_test0, y_train0, y_test0 = \
get_kth_imgs(N=50000,n=n,reCalc=False,resize=original_dim)
saved_G = '/home/ido/disk/KTH_G.bin'
saved_F = '/home/ido/disk/KTH_F.bin'
gg = []
ff = []
x_train0 = x_train0
x_train0[0],_ = synth2(224,H=0.3)
x_train0[1],_ = synth2(224,H=0.5)
for i in range(2):
x_train0[i]-=np.min(x_train0[i]*1.0)
x_train0[i]=x_train0[i]/np.max(x_train0[i])*256.0
print 'two first samples are fBm'
#plt.imshow(x_train0[0])
for i,im in enumerate(x_train0):
print i, '/', len(x_train0)
im = np.stack([im,im,im],axis=2).astype(np.float32) # grayscale
G0, F0 = texture.synTexture(im,onlyGram = True)
gg.append(G0)
ff.append(F0)
pickle.dump(gg,open(saved_G,'w'))
pickle.dump(ff,open(saved_F,'w'))
def prepare_data(self,idxs=None):
x_train0, x_test0, y_train0, y_test0 = \
get_kth_imgs(N=50000,n=self.n,reCalc=False,resize=self.original_dim)
x_train0[0],_ = synth2(224,H=0.3)
x_train0[1],_ = synth2(224,H=0.5)
for i in range(2):
x_train0[i]-=np.min(x_train0[i]*1.0)
x_train0[i]=x_train0[i]/np.max(x_train0[i])*256.0
print 'two first samples are fBm'
if idxs is not None:
return x_train0[idxs], x_test0, y_train0[idxs], y_test0
else:
return x_train0, x_test0, y_train0, y_test0
# train the VAE on MNIST digits
original_dim = 32 * 32 # 784
original_dim2 = int(np.sqrt(original_dim))
(x_train0, y_train0), (x_test0, y_test0) = mnist.load_data()
n = 32
use_fbms = False
batch_size = 100
if use_fbms:
print('USING fBms')
x_train0, x_test0, y_train0, y_test0 = \
generate_2d_fbms(N=50000,n=n,reCalc=False,resize=original_dim)
else:
print('using real imgs')
x_train0, x_test0, y_train0, y_test0 = \
get_kth_imgs(N=50000,n=n,reCalc=False,resize=original_dim)
localphase = LocalPhase(x_train0,
x_test0,
y_train0,
y_test0,
recalc_stats=True)
y_test0 = localphase.agg_y_test
y_train0 = localphase.agg_y_train
maxlen = (len(x_train0) / batch_size) * batch_size
x_train0 = x_train0[:maxlen]
y_train0 = y_train0[:maxlen]
maxlen = (len(x_test0) / batch_size) * batch_size
x_test0 = x_test0[:maxlen]
y_test0 = y_test0[:maxlen]
############################################################