def load(backend="numpy",binary = False, want_dense = False):
s=60000
T=np.zeros((s,10))
data_=open(nn.nas_address()+'/PSI-Share-no-backup/Ali/Dataset/MNIST/train-images.idx3-ubyte','r')
data=data_.read()
data_=np.fromstring(data[16:], dtype='uint8')
X=np.reshape(data_,(s,784))/255.0
data_=open(nn.nas_address()+'/PSI-Share-no-backup/Ali/Dataset/MNIST/train-labels.idx1-ubyte','r')
data=data_.read()
T_train_labels = np.fromstring(data[8:], dtype='uint8')
for n in range(s):
T[n,T_train_labels[n]]=1
s_test=10000
X_test=np.zeros((s_test,784))
T_test=np.zeros((s_test,10))
data_=open(nn.nas_address()+'/PSI-Share-no-backup/Ali/Dataset/MNIST/t10k-images.idx3-ubyte','r')
data=data_.read()
data_=np.fromstring(data[16:], dtype='uint8')
X_test=np.reshape(data_,(s_test,784))/255.0
data_=open(nn.nas_address()+'/PSI-Share-no-backup/Ali/Dataset/MNIST/t10k-labels.idx1-ubyte','r')
data=data_.read()
T_labels = np.fromstring(data[8:], dtype='uint8')
T_labels = T_labels.astype("float32")
for n in range(s_test):
T_test[n,T_labels[n]]=1
if binary:
X[X>.5]=1
X[X<.5]=0
X_test[X_test>.5]=1.0
X_test[X_test<.5]=0.0
if want_dense==False:
X = X.reshape(60000,1,28,28)
X_test = X_test.reshape(10000,1,28,28)
if backend=="numpy": X=nn.array(X);T=nn.array(T);X_test=nn.array(X_test);T_test=nn.array(T_test);T_train_labels=nn.array(T_train_labels);T_labels=nn.array(T_labels)
if backend=="gnumpy": X=nn.garray(X);T=nn.garray(T);X_test=nn.garray(X_test);T_test=nn.garray(T_test);T_train_labels=nn.garray(T_train_labels);T_labels=nn.garray(T_labels)
# print X.dtype,T.dtype
return X,T,X_test,T_test,T_train_labels,T_labels
def load(backend="numpy", bias=None, raw=False):
f = scipy.io.loadmat(
nn.nas_address() +
"/PSI-Share-no-backup/Ali/Dataset/Frey/frey_rawface.mat")
X = f['ff'].T
X_mean = X.mean(axis=1)
if not raw: X_std = (X.var(1) + 10)**.5
if not raw: X = (X - X_mean[:, np.newaxis]) / X_std[:, np.newaxis]
else: X = (X - X_mean[:, np.newaxis])
# M = X.mean(axis=0)
# X -= M
def cov(X):
X_mean = X.mean(axis=0)
X -= X_mean
return np.dot(X.T, X) / (1.0 * X.shape[0] - 1)
if not raw and bias:
sigma = cov(X)
u, s, v = np.linalg.svd(sigma)
P = np.dot(np.dot(u, np.diag(np.sqrt(1. / (s + bias)))), u.T)
X = np.dot(X, P)
X = X.reshape(1965, 1, 28, 20)
X = X[:, :, 3:-5, :]
if backend == "numpy":
return X
else:
return nn.garray(X)
def extract_patch(backend="numpy",num_patch=10000,size=7,want_desne=True):
img = Natural.load()
assert(img.ndim)==4
# img = img.reshape(50000,3,32,32)
# print img.max()
# nn.show_images(img[:9,:,:,:],(3,3))
X = np.zeros((num_patch,img.shape[1],size,size))
# extract random patches
for index in xrange(num_patch):
if index%100000==0: print index
x = random.randint(0,img.shape[2]-size)
y = random.randint(0,img.shape[3]-size)
patch = img[index%img.shape[0],:,x:x+size,y:y+size]
X[index,:,:,:] = patch
if want_desne:
X = X.reshape(num_patch,size*size)
if backend=="numpy":
return X
if backend=="gnumpy":
return nn.garray(X)
def extract_patch(img, backend="numpy", num_patch=10000, size=7):
# work_address = os.environ["WORK"]
# assert whitened == True
# if whitened:
# f=np.load(work_address+"./Dataset/CIFAR10/cifar_bias_.1.npz")
# img=f['X'];T=f['T'];X_test=f['X_test'];T_test=f['T_test'];T_train_labels=f['T_train_labels'];T_labels=f['T_labels']
# else:
# img,T,X_test,T_test,T_train_labels,T_labels = dataset.load_cifar10(raw=True)
assert (img.ndim) == 4
# img = img.reshape(50000,3,32,32)
# print img.max()
# nn.show_images(img[:9,:,:,:],(3,3))
X = np.zeros((num_patch, img.shape[1], size, size))
# extract random patches
for index in xrange(num_patch):
if index % 10000 == 0: print index
x = random.randint(0, img.shape[2] - size)
y = random.randint(0, img.shape[3] - size)
patch = img[index % img.shape[0], :, x:x + size, y:y + size]
X[index, :, :, :] = patch
if backend == "numpy":
return X
if backend == "gnumpy":
return nn.garray(X)
def extract_patch(backend="numpy", num_patch=10000, size=7, want_desne=True):
img = Natural.load()
assert (img.ndim) == 4
# img = img.reshape(50000,3,32,32)
# print img.max()
# nn.show_images(img[:9,:,:,:],(3,3))
X = np.zeros((num_patch, img.shape[1], size, size))
# extract random patches
for index in xrange(num_patch):
if index % 100000 == 0:
print index
x = random.randint(0, img.shape[2] - size)
y = random.randint(0, img.shape[3] - size)
patch = img[index % img.shape[0], :, x : x + size, y : y + size]
X[index, :, :, :] = patch
if want_desne:
X = X.reshape(num_patch, size * size)
if backend == "numpy":
return X
if backend == "gnumpy":
return nn.garray(X)
def load(backend="numpy", bias=None,raw = False):
f = scipy.io.loadmat(nn.nas_address()+"/PSI-Share-no-backup/Ali/Dataset/Frey/frey_rawface.mat")
X = f['ff'].T
X_mean=X.mean(axis=1)
if not raw: X_std = (X.var(1)+10)**.5
if not raw: X = (X-X_mean[:, np.newaxis])/X_std[:, np.newaxis]
else: X = (X-X_mean[:, np.newaxis])
# M = X.mean(axis=0)
# X -= M
def cov(X):
X_mean = X.mean(axis=0)
X -= X_mean
return np.dot(X.T,X)/(1.0*X.shape[0]-1)
if not raw and bias:
sigma = cov(X)
u,s,v=np.linalg.svd(sigma)
P = np.dot(np.dot(u,np.diag(np.sqrt(1./(s+bias)))),u.T)
X=np.dot(X,P)
X = X.reshape(1965,1,28,20)
X = X[:,:,3:-5,:]
if backend=="numpy":
return X
else:
return nn.garray(X)
def extract_patch(img,backend="numpy",num_patch=10000,size=7):
# work_address = os.environ["WORK"]
# assert whitened == True
# if whitened:
# f=np.load(work_address+"./Dataset/CIFAR10/cifar_bias_.1.npz")
# img=f['X'];T=f['T'];X_test=f['X_test'];T_test=f['T_test'];T_train_labels=f['T_train_labels'];T_labels=f['T_labels']
# else:
# img,T,X_test,T_test,T_train_labels,T_labels = dataset.load_cifar10(raw=True)
assert(img.ndim)==4
# img = img.reshape(50000,3,32,32)
# print img.max()
# nn.show_images(img[:9,:,:,:],(3,3))
X = np.zeros((num_patch,img.shape[1],size,size))
# extract random patches
for index in xrange(num_patch):
if index%10000==0: print index
x = random.randint(0,img.shape[2]-size)
y = random.randint(0,img.shape[3]-size)
patch = img[index%img.shape[0],:,x:x+size,y:y+size]
X[index,:,:,:] = patch
if backend=="numpy":
return X
if backend=="gnumpy":
return nn.garray(X)
def load_cifar10_filter(backend="numpy", size = 5):
assert backend == "gnumpy"
X,T,X_test,T_test,T_train_labels,T_labels=load_cifar10(raw=True)
X_mean=X.mean(axis=1)
X_test_mean=X_test.mean(axis=1)
X_std = (X.var(1)+10)**.5
X_test_std = (X_test.var(1)+10)**.5
X = (X-X_mean[:, np.newaxis])/X_std[:, np.newaxis]
X_test = (X_test-X_test_mean[:, np.newaxis])/X_test_std[:, np.newaxis]
X = X.reshape(50000,3,32,32)
X_test = X_test.reshape(10000,3,32,32)
filter = np.zeros((3,3,size,size))
T=nn.garray(T);T_test=nn.garray(T_test);T_train_labels=nn.garray(T_train_labels);T_labels=nn.garray(T_labels)
for i in xrange(3):
b = np.ones((size,size))*(1.0/size**2)
a = np.zeros((size,size))
a[(size-1)/2,(size-1)/2]=1
filter[i,i,:,:] = a-b
# X_filter = nn.ConvUp(X, filter, moduleStride = 1, paddingStart = (size-1)/2)
# X_filter_test = nn.ConvUp(X_test, filter, moduleStride = 1, paddingStart = (size-1)/2)
X_filter = nn.GnumpyBackend.zeros((50000,3,32,32))
for i in range(250):
X_filter[i*200:(i+1)*200,:,:,:] = nn.ConvUp(nn.garray(X[i*200:(i+1)*200,:,:,:]), filter, moduleStride = 1, paddingStart = (size-1)/2)
X_filter_test = nn.GnumpyBackend.zeros((10000,3,32,32))
for i in range(50):
X_filter_test[i*200:(i+1)*200,:,:,:] = nn.ConvUp(nn.garray(X_test[i*200:(i+1)*200,:,:,:]), filter, moduleStride = 1, paddingStart = (size-1)/2)
if backend=="gnumpy":
return X_filter,T,X_filter_test,T_test,T_train_labels,T_labels
def load_whiten(backend="numpy",bias = .1):
X,T,X_test,T_test,T_train_labels,T_labels=CIFAR10.load(want_mean=False,want_dense=True)
#normalize for contrast
X_mean=X.mean(axis=1)
X_test_mean=X_test.mean(axis=1)
X_std = X.std(1)
X_test_std = X_test.std(1)
X = (X-X_mean[:, np.newaxis])/X_std[:, np.newaxis]
X_test = (X_test-X_test_mean[:, np.newaxis])/X_test_std[:, np.newaxis]
#covariance
def cov(X):
X_mean = X.mean(axis=0)
X -= X_mean
return np.dot(X.T,X)/(1.0*X.shape[0]-1)
#whiten
X_mean = X.mean(axis=0)
X -= X_mean
X_test -= X_mean
sigma = cov(X)
u,s,v=np.linalg.svd(sigma)
P = np.dot(np.dot(u,np.diag(np.sqrt(1./(s+bias)))),u.T)
X=np.dot(X,P)
X_test=np.dot(X_test,P)
X = X.reshape(50000,3,32,32)
X_test = X_test.reshape(10000,3,32,32)
if backend=="numpy":
return X,T,X_test,T_test,T_train_labels,T_labels
if backend=="gnumpy":
return nn.garray(X),nn.garray(T),nn.garray(X_test),nn.garray(T_test),nn.garray(T_train_labels),nn.garray(T_labels)
def load_cifar10_filter(backend="numpy", size=5):
assert backend == "gnumpy"
X, T, X_test, T_test, T_train_labels, T_labels = load_cifar10(raw=True)
X_mean = X.mean(axis=1)
X_test_mean = X_test.mean(axis=1)
X_std = (X.var(1) + 10)**.5
X_test_std = (X_test.var(1) + 10)**.5
X = (X - X_mean[:, np.newaxis]) / X_std[:, np.newaxis]
X_test = (X_test - X_test_mean[:, np.newaxis]) / X_test_std[:,
np.newaxis]
X = X.reshape(50000, 3, 32, 32)
X_test = X_test.reshape(10000, 3, 32, 32)
filter = np.zeros((3, 3, size, size))
T = nn.garray(T)
T_test = nn.garray(T_test)
T_train_labels = nn.garray(T_train_labels)
T_labels = nn.garray(T_labels)
for i in xrange(3):
b = np.ones((size, size)) * (1.0 / size**2)
a = np.zeros((size, size))
a[(size - 1) / 2, (size - 1) / 2] = 1
filter[i, i, :, :] = a - b
# X_filter = nn.ConvUp(X, filter, moduleStride = 1, paddingStart = (size-1)/2)
# X_filter_test = nn.ConvUp(X_test, filter, moduleStride = 1, paddingStart = (size-1)/2)
X_filter = nn.GnumpyBackend.zeros((50000, 3, 32, 32))
for i in range(250):
X_filter[i * 200:(i + 1) * 200, :, :, :] = nn.ConvUp(
nn.garray(X[i * 200:(i + 1) * 200, :, :, :]),
filter,
moduleStride=1,
paddingStart=(size - 1) / 2)
X_filter_test = nn.GnumpyBackend.zeros((10000, 3, 32, 32))
for i in range(50):
X_filter_test[i * 200:(i + 1) * 200, :, :, :] = nn.ConvUp(
nn.garray(X_test[i * 200:(i + 1) * 200, :, :, :]),
filter,
moduleStride=1,
paddingStart=(size - 1) / 2)
if backend == "gnumpy":
return X_filter, T, X_filter_test, T_test, T_train_labels, T_labels
def load_whiten(backend="numpy", bias=.1):
X, T, X_test, T_test, T_train_labels, T_labels = CIFAR10.load(
want_mean=False, want_dense=True)
#normalize for contrast
X_mean = X.mean(axis=1)
X_test_mean = X_test.mean(axis=1)
X_std = X.std(1)
X_test_std = X_test.std(1)
X = (X - X_mean[:, np.newaxis]) / X_std[:, np.newaxis]
X_test = (X_test - X_test_mean[:, np.newaxis]) / X_test_std[:,
np.newaxis]
#covariance
def cov(X):
X_mean = X.mean(axis=0)
X -= X_mean
return np.dot(X.T, X) / (1.0 * X.shape[0] - 1)
#whiten
X_mean = X.mean(axis=0)
X -= X_mean
X_test -= X_mean
sigma = cov(X)
u, s, v = np.linalg.svd(sigma)
P = np.dot(np.dot(u, np.diag(np.sqrt(1. / (s + bias)))), u.T)
X = np.dot(X, P)
X_test = np.dot(X_test, P)
X = X.reshape(50000, 3, 32, 32)
X_test = X_test.reshape(10000, 3, 32, 32)
if backend == "numpy":
return X, T, X_test, T_test, T_train_labels, T_labels
if backend == "gnumpy":
return nn.garray(X), nn.garray(T), nn.garray(X_test), nn.garray(
T_test), nn.garray(T_train_labels), nn.garray(T_labels)
def load_cifar10_adam_patch(num_patch,
size,
backend="numpy",
want_dense=True):
img, T, X_test, T_test, T_train_labels, T_labels = CIFAR10.load(
want_mean=False, want_dense=True)
img = img.reshape(50000, 3, 32, 32)
# print img.max()
# nn.show_images(img[:9,:,:,:],(3,3))
X = np.zeros((num_patch, 3, size, size))
# extract random patches
# for index in xrange(num_patch):
# x = random.randint(0,31-size)
# y = random.randint(0,31-size)
# patch = img[index%50000,:,x:x+size,y:y+size]
# X[index,:,:,:] = patch
for index in xrange(num_patch):
if index % 10000 == 0: print index
x = random.randint(0, img.shape[2] - size)
y = random.randint(0, img.shape[3] - size)
patch = img[index % img.shape[0], :, x:x + size, y:y + size]
X[index, :, :, :] = patch
X = X.reshape(num_patch, -1)
#normalize for contrast
X_mean = X.mean(axis=1)
print X_mean.shape
# X_std = (X.std(1)+10)**.5
# X_std = X.std(1)
X_std = (X.var(1) + 10)**.5
X = (X - X_mean[:, np.newaxis]) / X_std[:, np.newaxis]
#covariance
def cov(X):
X_mean = X.mean(axis=0)
X -= X_mean
return np.dot(X.T, X) / (1.0 * X.shape[0] - 1)
#whiten
sigma = cov(X) + .1 * np.identity(X.shape[1])
M = X.mean(axis=0)
X -= M
u, s, v = np.linalg.svd(sigma)
P = np.dot(np.dot(u, np.diag(np.sqrt(1. / s))), u.T)
X = np.dot(X, P)
# X=nn.garray(X)
# print X.shape
# np.savez("cifar10_adam", X = X)
# print X.min(0)
# print (X**2).sum(1)[:100]
if not want_dense:
X = X.reshape(num_patch, 3, size, size)
if backend == "numpy":
return X, M, P
if backend == "gnumpy":
return nn.garray(X)
def load(backend="numpy", binary=False, want_dense=False):
s = 60000
T = np.zeros((s, 10))
data_ = open(
nn.nas_address() +
'/PSI-Share-no-backup/Ali/Dataset/MNIST/train-images.idx3-ubyte',
'r')
data = data_.read()
data_ = np.fromstring(data[16:], dtype='uint8')
X = np.reshape(data_, (s, 784)) / 255.0
data_ = open(
nn.nas_address() +
'/PSI-Share-no-backup/Ali/Dataset/MNIST/train-labels.idx1-ubyte',
'r')
data = data_.read()
T_train_labels = np.fromstring(data[8:], dtype='uint8')
for n in range(s):
T[n, T_train_labels[n]] = 1
s_test = 10000
X_test = np.zeros((s_test, 784))
T_test = np.zeros((s_test, 10))
data_ = open(
nn.nas_address() +
'/PSI-Share-no-backup/Ali/Dataset/MNIST/t10k-images.idx3-ubyte',
'r')
data = data_.read()
data_ = np.fromstring(data[16:], dtype='uint8')
X_test = np.reshape(data_, (s_test, 784)) / 255.0
data_ = open(
nn.nas_address() +
'/PSI-Share-no-backup/Ali/Dataset/MNIST/t10k-labels.idx1-ubyte',
'r')
data = data_.read()
T_labels = np.fromstring(data[8:], dtype='uint8')
T_labels = T_labels.astype("float32")
for n in range(s_test):
T_test[n, T_labels[n]] = 1
if binary:
X[X > .5] = 1
X[X < .5] = 0
X_test[X_test > .5] = 1.0
X_test[X_test < .5] = 0.0
if want_dense == False:
X = X.reshape(60000, 1, 28, 28)
X_test = X_test.reshape(10000, 1, 28, 28)
if backend == "numpy":
X = nn.array(X)
T = nn.array(T)
X_test = nn.array(X_test)
T_test = nn.array(T_test)
T_train_labels = nn.array(T_train_labels)
T_labels = nn.array(T_labels)
if backend == "gnumpy":
X = nn.garray(X)
T = nn.garray(T)
X_test = nn.garray(X_test)
T_test = nn.garray(T_test)
T_train_labels = nn.garray(T_train_labels)
T_labels = nn.garray(T_labels)
# print X.dtype,T.dtype
return X, T, X_test, T_test, T_train_labels, T_labels
def load_cifar10_adam_patch(num_patch,size,backend="numpy",want_dense = True):
img,T,X_test,T_test,T_train_labels,T_labels=CIFAR10.load(want_mean = False, want_dense = True)
img = img.reshape(50000,3,32,32)
# print img.max()
# nn.show_images(img[:9,:,:,:],(3,3))
X = np.zeros((num_patch,3,size,size))
# extract random patches
# for index in xrange(num_patch):
# x = random.randint(0,31-size)
# y = random.randint(0,31-size)
# patch = img[index%50000,:,x:x+size,y:y+size]
# X[index,:,:,:] = patch
for index in xrange(num_patch):
if index%10000==0: print index
x = random.randint(0,img.shape[2]-size)
y = random.randint(0,img.shape[3]-size)
patch = img[index%img.shape[0],:,x:x+size,y:y+size]
X[index,:,:,:] = patch
X = X.reshape(num_patch,-1)
#normalize for contrast
X_mean=X.mean(axis=1)
print X_mean.shape
# X_std = (X.std(1)+10)**.5
# X_std = X.std(1)
X_std = (X.var(1)+10)**.5
X = (X-X_mean[:, np.newaxis])/X_std[:, np.newaxis]
#covariance
def cov(X):
X_mean = X.mean(axis=0)
X -= X_mean
return np.dot(X.T,X)/(1.0*X.shape[0]-1)
#whiten
sigma = cov(X)+.1*np.identity(X.shape[1])
M = X.mean(axis=0)
X -= M
u,s,v=np.linalg.svd(sigma)
P = np.dot(np.dot(u,np.diag(np.sqrt(1./s))),u.T)
X=np.dot(X,P)
# X=nn.garray(X)
# print X.shape
# np.savez("cifar10_adam", X = X)
# print X.min(0)
# print (X**2).sum(1)[:100]
if not want_dense:
X = X.reshape(num_patch,3,size,size)
if backend=="numpy":
return X,M,P
if backend=="gnumpy":
return nn.garray(X)
