def __init__(self, config):
"""Constructor.
"""
self.config = config
idxs = []
with open(self.config['path'] + '/seq_ref_clean.txt',
'r',
encoding="latin-1") as f:
for l, line in enumerate(f):
if l >= 4:
idxs.append([int(j) for j in line.split(' ')])
idxs[-1][1] = 4371
X = self.config['pad_value'] * np.ones(
(len(idxs), np.max(np.diff(np.array(idxs), axis=1)) + 1,
self.config['size'], self.config['size'], 3),
dtype=self.config['dtype'])
y = np.zeros((len(idxs), ), dtype='int32')
for i, idx in enumerate(idxs):
for j, q in enumerate(range(idx[0], idx[1] + 1)):
img = imread(self.config['path'] + '/img_clean/img' +
'{:04d}'.format(q) + '.jpg')
img = imresize(img, (self.config['size'], self.config['size']))
X[i, j, :, :, :] = img
y[i] = i
X /= 255.
# zca
if self.config['zca_epsilon'] is not None:
self.ZCA = ZCAWhitening(epsilon=self.config['zca_epsilon'])
X = self.ZCA.fit_transform(X)
if self.config['imagenet_preprocessing'] is True:
X = np.pad(X, ((0, 0), (0, 0), (0, 32 - self.config['size']),
(0, 32 - self.config['size']), (0, 0)), 'constant')
for i in range(len(X)):
for t in range(len(X[0])):
X[i, t] = imresize(X[i, t], (32, 32))
X = preprocess_input(X, data_format='channels_last')
Y = to_categorical(y, len(X))
self.X = X
self.Y = Y
print(self.X.shape, self.Y.shape)
class COIL100Dataset(MNISTRotatedDataset):
def __init__(self, config):
"""Constructor.
"""
self.config = config
X = np.zeros((100, 72, self.config['size'], self.config['size'], 3),
dtype=self.config['dtype'])
y = np.zeros((100, ), dtype='int32')
for i in range(len(X)):
for j, t in enumerate(range(0, 360, 5)):
img = imread(self.config['path'] + '/obj' + str(i + 1) + '__' +
str(t) + '.png')
img = imresize(img, (self.config['size'], self.config['size']))
X[i, j, :, :, :] = img
y[i] = i
X /= 255.
# zca
if self.config['zca_epsilon'] is not None:
self.ZCA = ZCAWhitening(epsilon=self.config['zca_epsilon'])
X = self.ZCA.fit_transform(X)
if self.config['imagenet_preprocessing'] is True:
X = np.pad(X, ((0, 0), (0, 0), (0, 32 - self.config['size']),
(0, 32 - self.config['size']), (0, 0)), 'constant')
for i in range(len(X)):
for t in range(len(X[0])):
X[i, t] = imresize(X[i, t], (32, 32))
X = preprocess_input(X, data_format='channels_last')
Y = to_categorical(y, len(X))
self.X = X
self.Y = Y
print(self.X.shape, self.Y.shape)
def generate_labeled_data(self, num_labeled):
if num_labeled is None:
num_labeled = self.X_train[1]
ind_train = list(
range(0, 64 - np.mod(64, 64 // num_labeled), 64 // num_labeled))
ind_val = list(
set(
range((64 // num_labeled) // 2, 64 -
np.mod(64, 64 // num_labeled), 64 // num_labeled)) -
set(ind_train))
ind_test = list(
set(range(self.X.shape[1])) - set(ind_train) - set(ind_val))
self.X_train_labeled = self.X[:, ind_train]
self.Y_train_labeled = self.Y
self.X_val_labeled = self.X[:, ind_val]
self.Y_val_labeled = self.Y
self.X_test = self.X[:, ind_test]
self.Y_test = self.Y
print(self.X_train_labeled.shape, self.Y_train_labeled.shape,
self.X_val_labeled.shape, self.Y_val_labeled.shape,
self.X_test.shape, self.Y_test.shape)
def __init__(self, config):
"""Constructor.
"""
self.config = config
X = np.zeros((100, 72, self.config['size'], self.config['size'], 3),
dtype=self.config['dtype'])
y = np.zeros((100, ), dtype='int32')
for i in range(len(X)):
for j, t in enumerate(range(0, 360, 5)):
img = imread(self.config['path'] + '/obj' + str(i + 1) + '__' +
str(t) + '.png')
img = imresize(img, (self.config['size'], self.config['size']))
X[i, j, :, :, :] = img
y[i] = i
X /= 255.
# zca
if self.config['zca_epsilon'] is not None:
self.ZCA = ZCAWhitening(epsilon=self.config['zca_epsilon'])
X = self.ZCA.fit_transform(X)
if self.config['imagenet_preprocessing'] is True:
X = np.pad(X, ((0, 0), (0, 0), (0, 32 - self.config['size']),
(0, 32 - self.config['size']), (0, 0)), 'constant')
for i in range(len(X)):
for t in range(len(X[0])):
X[i, t] = imresize(X[i, t], (32, 32))
X = preprocess_input(X, data_format='channels_last')
Y = to_categorical(y, len(X))
self.X = X
self.Y = Y
print(self.X.shape, self.Y.shape)
class MNISTRotatedDataset():
def __init__(self, config):
"""Constructor.
"""
self.config = config
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train = X_train.reshape(-1, 28, 28, 1)
X_test = X_test.reshape(-1, 28, 28, 1)
X_train = X_train.astype(self.config['dtype'])
X_test = X_test.astype(self.config['dtype'])
X_train_rot = np.zeros((X_train.shape[0], self.config['rotations'], self.config['size'], self.config['size'], 1))
X_test_rot = np.zeros((X_test.shape[0], self.config['rotations'], self.config['size'], self.config['size'], 1))
for i in range(len(X_train)):
img = Image.fromarray(np.reshape(X_train[i, :, :, :], (28, 28)))
img = img.convert('L')
phase = 360. * np.random.rand()
for t in range(self.config['rotations']):
_img = img.rotate(phase + t * 360. / self.config['rotations'], Image.BILINEAR).resize((28, 28), Image.BILINEAR).getdata()
_img = np.reshape(np.array(_img, dtype=np.uint8), (28, 28, 1))[:self.config['size'], :self.config['size'], :]
X_train_rot[i, t, :, :, :] = _img.astype(dtype=self.config['dtype'])
for i in range(len(X_test)):
img = Image.fromarray(np.reshape(X_test[i, :, :, :], (28, 28)))
img = img.convert('L')
phase = 360. * np.random.rand()
for t in range(self.config['rotations']):
_img = img.rotate(phase + t * 360. / self.config['rotations'], Image.BILINEAR).resize((28, 28), Image.BILINEAR).getdata()
_img = np.reshape(np.array(_img, dtype=np.uint8), (28, 28, 1))[:self.config['size'], :self.config['size'], :]
X_test_rot[i, t, :, :, :] = _img.astype(dtype=self.config['dtype'])
X_train_rot /= 255.
X_test_rot /= 255.
# zca
if self.config['zca_epsilon'] is not None:
self.ZCA = ZCAWhitening(epsilon=self.config['zca_epsilon'])
X_train_rot = self.ZCA.fit_transform(X_train_rot)
X_test_rot = self.ZCA.transform(X_test_rot)
if self.config['imagenet_preprocessing'] is True:
X_train_rot = np.pad(X_train_rot, ((0, 0), (0, 0), (0, 32 - self.config['size']), (0, 32 - self.config['size']), (0, 0)), 'constant')
X_test_rot = np.pad(X_test_rot, ((0, 0), (0, 0), (0, 32 - self.config['size']), (0, 32 - self.config['size']), (0, 0)), 'constant')
X_train_rot = np.repeat(X_train_rot, 3, axis=-1)
X_test_rot = np.repeat(X_test_rot, 3, axis=-1)
for i in range(len(X_train_rot)):
for t in range(len(X_train_rot[0])):
X_train_rot[i, t] = imresize(X_train_rot[i, t], (32, 32))
for i in range(len(X_test_rot)):
for t in range(len(X_test_rot[0])):
X_test_rot[i, t] = imresize(X_test_rot[i, t], (32, 32))
X_train_rot = preprocess_input(X_train_rot, data_format='channels_last')
X_test_rot = preprocess_input(X_test_rot, data_format='channels_last')
nb_classes = 10
Y_train = to_categorical(y_train, nb_classes)
Y_test = to_categorical(y_test, nb_classes)
self.X_train = X_train_rot
self.Y_train = Y_train
self.X_test = X_test_rot
self.Y_test = Y_test
print(self.X_train.shape, self.Y_train.shape, self.X_test.shape, self.Y_test.shape)
def generate_labeled_data(self, num_labeled=None):
if num_labeled is None:
num_labeled = round(0.9 * len(self.X_train))
while True:
perm = np.random.permutation(len(self.X_train))
self.X_train_labeled = self.X_train[perm[:num_labeled]]
self.Y_train_labeled = self.Y_train[perm[:num_labeled]]
self.X_val_labeled = self.X_train[perm[num_labeled:round(1.1111 * num_labeled)]]
self.Y_val_labeled = self.Y_train[perm[num_labeled:round(1.1111 * num_labeled)]]
if np.all(np.any(self.Y_train_labeled, axis=0)) and np.all(np.any(self.Y_val_labeled, axis=0)):
break
print(self.X_train_labeled.shape, self.Y_train_labeled.shape, self.X_val_labeled.shape, self.Y_val_labeled.shape, self.X_test.shape, self.Y_test.shape)
class CIFAR10ScannedDataset(MNISTRotatedDataset):
def __init__(self, config):
"""Constructor.
"""
self.config = config
(X_train, y_train), (X_test, y_test) = cifar10.load_data()
X_train = X_train.reshape(-1, 32, 32, 3)
X_test = X_test.reshape(-1, 32, 32, 3)
X_train = X_train.astype(self.config['dtype'])
X_test = X_test.astype(self.config['dtype'])
N = ((32 - self.config['size']) // self.config['stride']) + 1
X_train_scan = np.zeros((X_train.shape[0], N**2, self.config['size'],
self.config['size'], 3))
X_test_scan = np.zeros((X_test.shape[0], N**2, self.config['size'],
self.config['size'], 3))
for i in range(len(X_train)):
for t1 in range(N):
for t2 in (range(N) if
(np.mod(t1, 2) == 0) else range(N - 1, -1, -1)):
X_train_scan[i, t1 * N +
(t2 if (np.mod(t1, 2) == 0) else
(N - 1 - t2)), :, :, :] = X_train[i, (
t2 * self.config['stride']):(
t2 * self.config['stride'] +
self.config['size']), (
t1 * self.config['stride']):(
t1 * self.config['stride'] +
self.config['size']), :]
for i in range(len(X_test)):
for t1 in range(N):
for t2 in (range(N) if
(np.mod(t1, 2) == 0) else range(N - 1, -1, -1)):
X_test_scan[i,
t1 * N + (t2 if (np.mod(t1, 2) == 0) else
(N - 1 - t2)), :, :, :] = X_test[
i, (t2 * self.config['stride']):(
t2 * self.config['stride'] +
self.config['size']),
(t1 * self.config['stride']):(
t1 * self.config['stride'] +
self.config['size']), :]
X_train_scan /= 255.
X_test_scan /= 255.
# zca
if self.config['zca_epsilon'] is not None:
self.ZCA = ZCAWhitening(epsilon=self.config['zca_epsilon'])
X_train_scan = self.ZCA.fit_transform(X_train_scan)
X_test_scan = self.ZCA.transform(X_test_scan)
if self.config['imagenet_preprocessing'] is True:
X_train_scan = np.pad(X_train_scan,
((0, 0), (0, 0),
(0, 32 - self.config['size']),
(0, 32 - self.config['size']), (0, 0)),
'constant')
X_test_scan = np.pad(X_test_scan,
((0, 0), (0, 0),
(0, 32 - self.config['size']),
(0, 32 - self.config['size']), (0, 0)),
'constant')
for i in range(len(X_train_scan)):
for t in range(len(X_train_scan[0])):
X_train_scan[i, t] = imresize(X_train_scan[i, t], (32, 32))
for i in range(len(X_test_scan)):
for t in range(len(X_test_scan[0])):
X_test_scan[i, t] = imresize(X_test_scan[i, t], (32, 32))
X_train_scan = preprocess_input(X_train_scan,
data_format='channels_last')
X_test_scan = preprocess_input(X_test_scan,
data_format='channels_last')
nb_classes = 10
Y_train = to_categorical(y_train, nb_classes)
Y_test = to_categorical(y_test, nb_classes)
self.X_train = X_train_scan
self.Y_train = Y_train
self.X_test = X_test_scan
self.Y_test = Y_test
print(self.X_train.shape, self.Y_train.shape, self.X_test.shape,
self.Y_test.shape)