def get_stats(im):
def standardize(x):
x = np.array(x)
x[np.isnan(x)] = 0.0
x = x - np.min(x)
x = x / np.max(x)
return x
patch = standardize(im)
h, _ = hurst2d(patch, max_tau=7)
# estimate Gaussianity via kurtosis
kurt = kurtosis(patch.flatten())
coh = coherence(patch)
return dict({'H': h, 'Kurtosis': kurt, 'MeanCoh': np.mean(coh['logcoh'])})
def get_kurtsim_imgs(N=120,n=32,reCalc=False,resize=None):
#reCalc = True
#reCalc = False
fname = 'data_kurtsim.bin'
if os.path.exists(fname) and not reCalc:
print 'loading data from file'
Xtrain, Xtest, Ytrain, Ytest = pickle.load(open(fname,'r'))
else:
print 'generating data and saving'
np.random.seed(0)
X = []
stds = np.linspace(0.0,0.4,N)# [0.5]*N
feats=[]
im = cv2.imread('/home/ido/brodatz/1.1.01.tiff').astype(np.float32)
im = im[:224,:224,0]
im = im-np.min(im)
im0 = im/np.max(im)
#plt.imshow(im)
for k,s in enumerate(stds):
if not k % 20:
print 'done %f'%(1.0*k/N)
im = im0 + np.random.randn(224,224)*s
X.append(im)
h = hurst2d(im,max_tau=7)
kurt = kurtosis(im.flatten())
coh = coherence(im)
#print(kurt)
feats.append([h, kurt, np.mean(coh['logcoh']), np.std(coh['logcoh'])])
X0=np.array(X)
Y = np.array(feats)
#X=np.expand_dims(X0,2) # to make input of size (n,1) instead of just (n)
if resize is not None:
sz=int(np.sqrt(resize))
X0r = np.zeros([X0.shape[0],sz,sz])
for i in range(X0.shape[0]):
X0r[i,:,:] = imresize(X0[i,:,:],[sz,sz])
#print X0r[i,:,:]
#plt.imshow(X0r[i,:,:],interpolation='none',cmap='gray')
#plt.show()
X0=X0r
Xtrain, Xtest, Ytrain, Ytest = train_test_split(X0,Y,test_size=0.2,random_state=0)
pickle.dump([Xtrain,Xtest,Ytrain,Ytest],open(fname,'w'))
return Xtrain, Xtest, Ytrain, Ytest
def get_kth_imgs(N=10000,n=32,reCalc=False,resize=None):
#reCalc = True
#reCalc = False
if os.path.exists('data_kth.bin') and not reCalc:
print 'loading data from file'
Xtrain, Xtest, Ytrain, Ytest = pickle.load(open('data_kth.bin','r'))
else:
print 'loading data and saving'
X = []
H = []
names = []
path0 = '/home/ido/combined_model_classification/KTH-TIPS2-b/'
samples = ['wool', 'cotton', 'cracker']
path = ['/sample_', '/']
dirs = ['a','b','c','d']
files = []
for sample in samples:
for d in dirs:
path1 = path0+sample+path[0]+d+path[1]
files += [path1+f for f in listdir(path1)]
names += [sample+d for i in range(len(listdir(path1)))]
print 'tot files:',len(files)
p = n
samp_names=[]
try:
for i,(f,name) in enumerate(zip(files,names)):
print i,'name',name,'file',f
img = Image.open(f).convert('L')
(width, height) = img.size
img = np.array(list(img.getdata()))
img = img.reshape((height, width))
sz = [width, height]
for k in range(0,sz[0],p):
for l in range(0,sz[1],p):
patch = img[k:k+p,l:l+p]*1.0
if np.min(patch.shape) == 0:
continue
patch = patch-np.min(patch)
patch = patch/np.max(patch)
#print patch
if patch.shape[0]<p or patch.shape[1]<p:
continue
h,_ = hurst2d(patch,max_tau=7)
# estimate Gaussianity via kurtosis
kurt = kurtosis(patch.flatten())
coh = coherence(patch)
#print(kurt)
H.append([h, kurt, name, np.mean(coh['logcoh']), np.std(coh['logcoh'])])
samp_names.append(name)
X.append(patch)
if len(X)>=N:
print 'too many images'
raise GetOutOfLoop
except GetOutOfLoop:
pass
print 'max N',N,'number of patches',len(X)
Y = np.vstack(H)
X = np.array(X)
if resize is not None:
sz=int(np.sqrt(resize))
X0r = np.zeros([X.shape[0],sz,sz])
for i in range(X.shape[0]):
X0r[i,:,:] = imresize(X[i,:,:],[sz,sz])
#print X0r[i,:,:]
#plt.imshow(X0r[i,:,:],interpolation='none',cmap='gray')
#plt.show()
X=X0r
#X = np.expand_dims(np.array(X),3) # to make input of size (n,1) instead of just (n)
Xtrain, Xtest, Ytrain, Ytest = train_test_split(X,Y,test_size=0.2,random_state=0)
pickle.dump([Xtrain,Xtest,Ytrain,Ytest],open('data_kth.bin','w'))
return Xtrain, Xtest, Ytrain, Ytest
#rr[0]=0
#z_sample = np.array([[xi, yi, zi, rr[0], rr[1]]])
z_sample = np.array([rr])
x_decoded = generator.predict(z_sample)
x_decoded = np.reshape(x_decoded[0],[n,n])
"""
plt.imshow(x_decoded,interpolation='none')
plt.show()
plt.pause(0.1)
continue
#"""
#digit = x_decoded[0].reshape(digit_size, digit_size)
x_decoded = x_decoded-np.min(x_decoded)
x_decoded = x_decoded/np.max(x_decoded)
H_est, rsquare = hurst2d(x_decoded,max_tau=7)
Hs.append(H_est)
rsquares+=rsquare
tot+=1
if not i % 30:
spl[0].hold(False)
diffres = np.squeeze(np.reshape(x_decoded[1:]-x_decoded[:-1],[1,-1]))
_, bins, patches = spl[0].hist(diffres, 50, normed=1, facecolor='green', alpha=0.75)
y = mlab.normpdf( bins, np.mean(diffres), np.std(diffres))
spl[0].hold(True)
spl[0].plot(bins, y, 'r--', linewidth=1)
spl[0].set_title('diff, exp %d/%d'%(i,num_exps))
spl[1].hold(False)
spl[1].imshow(x_decoded,interpolation='none')
#plt.title('y %f, x %f, z %f'%(yi,xi,zi))