labels.append(label)
f.create_dataset('images', data=np.array(images))
f.create_dataset('labels', data=np.array(labels))
f.attrs['index_matrix'] = index_matrix
f.attrs['class_names'] = np.array(
aid.classes, dtype=h5py.special_dtype(vlen=str))
# load square aid
f = h5py.File(data_directory + '/aid' + str(resize_size[0]) + '.h5', 'r')
data = f['images'][()]
labels = f['labels'][()]
class_names = f.attrs['class_names']
f.close()
# Normalize data
data = utils.normalize(data)
data = data.reshape(data.shape[0], data.shape[1], resize_size[0],
resize_size[1])
# Datasets
dataset = utils.NumpyDataset(data, labels, transform=utils.NumpyToTensor())
# Run the experiments
for seed in seeds:
# Data loaders
logger.info('Split data with seed {}'.format(seed))
torch.manual_seed(seed)
np.random.seed(seed)
train_indices = []
val_indices = []
index_matrix = index_matrix.reshape(train_data_all.shape[2],
train_data_all.shape[3])
train_data_all = train_data_all.reshape(train_data_all.shape[0],
train_data_all.shape[1], -1)
train_labels_all = train_set.train_labels
test_set = datasets.CIFAR10(data_directory, train=False)
test_data = test_set.test_data.transpose(0, 3, 1, 2).astype(np.float32)
test_data = test_data.reshape(test_data.shape[0], test_data.shape[1],
-1)
test_labels = test_set.test_labels
index_matrix.unsqueeze_(0)
index_matrix.unsqueeze_(0)
# Normalize data
train_data_all = utils.normalize(train_data_all)
test_data = utils.normalize(test_data)
# Whitening data
logger.info('Computing whitening matrices...')
train_data_all_flat = train_data_all.reshape(train_data_all.shape[0], -1).T
test_data_flat = test_data.reshape(test_data.shape[0], -1).T
pca_all = utils.PCA(D=train_data_all_flat,
n_components=train_data_all_flat.shape[1])
logger.info('Whitening data...')
train_data_all_flat = pca_all.transform(D=train_data_all_flat,
whiten=True,
ZCA=True)
train_data_all = train_data_all_flat.T.reshape(train_data_all.shape[0:2] +
(-1, ))