def arxiv(root, ids, new_name=None):
"""Download paper from https://arxiv.org, the file format is pdf.
Data storage directory:
root = `/user/.../mydata`
`ids`.pdf data:
`root/arxiv/`ids`.pdf` or `root/arxiv/`new_name`.pdf`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/arxiv`,
root should be `/user/.../mydata`.
ids: str, arxiv paper id.
example:ids = '1605.09782' mean you want get paper links https://arxiv.org/abs/1605.09782.
new_name: str, default None. if not None, download file path is `root/arxiv/new_name.pdf`.
Returns:
Store the absolute path of the data directory, is `root/arxiv`.
"""
start = time.time()
assert gfile.isdir(root), '`root` should be directory.'
assert isinstance(ids, str), '`ids` type should be str.'
if new_name is None:
task_path = gfile.path_join(root, 'arxiv', ids + '.pdf')
else:
task_path = gfile.path_join(root, 'arxiv', new_name + '.pdf')
gfile.makedirs(gfile.path_join(root, 'arxiv'))
gfile.remove(task_path)
url = 'https://arxiv.org/pdf/' + str(ids) + '.pdf'
rq.files(url, task_path)
print('arxiv paper download completed, run time %d min %.2f sec' % divmod(
(time.time() - start), 60))
return task_path
def mnist_fashion(root):
"""A dataset of Zalando's article images consisting of fashion products.
Fashion mnist datasets is a drop-in replacement of the original MNIST dataset
from https://github.com/zalandoresearch/fashion-mnist.
Each sample is an gray image (in 3D NDArray) with shape (28, 28, 1).
Attention: if exist dirs `root/mnist_fashion`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
mnist_fashion data:
`root/mnist_fashion/train/0/xx.png`
`root/mnist_fashion/train/2/xx.png`
`root/mnist_fashion/train/6/xx.png`
`root/mnist_fashion/test/0/xx.png`
`root/mnist_fashion/test/2/xx.png`
`root/mnist_fashion/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/mnist_fashion`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/mnist_fashion`.
"""
start = time.time()
task_path = assert_dirs(root, 'mnist_fashion')
url_list = ['http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz',
'http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz',
'http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz',
'http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz']
for url in url_list:
rq.files(url, gfile.path_join(task_path, url.split('/')[-1]))
with gzip.open(gfile.path_join(task_path, 'train-labels-idx1-ubyte.gz'), 'rb') as lbpath:
train_label = np.frombuffer(lbpath.read(), np.uint8, offset=8)
with gzip.open(gfile.path_join(task_path, 'train-images-idx3-ubyte.gz'), 'rb') as imgpath:
train = np.frombuffer(imgpath.read(), np.uint8, offset=16).reshape(len(train_label), 28, 28)
with gzip.open(gfile.path_join(task_path, 't10k-labels-idx1-ubyte.gz'), 'rb') as lbpath:
test_label = np.frombuffer(lbpath.read(), np.uint8, offset=8)
with gzip.open(gfile.path_join(task_path, 't10k-images-idx3-ubyte.gz'), 'rb') as imgpath:
test = np.frombuffer(imgpath.read(), np.uint8, offset=16).reshape(len(test_label), 28, 28)
for i in set(train_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for i in set(test_label):
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for idx in range(train.shape[0]):
save_image(gfile.path_join(task_path, 'train', str(train_label[idx]), str(idx)+'.png'),
array_to_image(train[idx].reshape(28, 28, 1)))
for idx in range(test.shape[0]):
save_image(gfile.path_join(task_path, 'test', str(test_label[idx]), str(idx)+'.png'),
array_to_image(test[idx].reshape(28, 28, 1)))
for url in url_list:
gfile.remove(gfile.path_join(task_path, url.split('/')[-1]))
print('mnist_fashion dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def cifar100(root, fine_label=True):
"""CIFAR100 image classification dataset from https://www.cs.toronto.edu/~kriz/cifar.html
Each sample is an image (in 3D NDArray) with shape (32, 32, 3).
Attention: if exist dirs `root/cifar100`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
cifar100 data:
`root/cifar100/train/0/xx.png`
`root/cifar100/train/2/xx.png`
`root/cifar100/train/6/xx.png`
`root/cifar100/test/0/xx.png`
`root/cifar100/test/2/xx.png`
`root/cifar100/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/cifar100`,
root should be `/user/.../mydata`.
fine_label: bool, default False.
Whether to load the fine-grained (100 classes) or
coarse-grained (20 super-classes) labels.
Returns:
Store the absolute path of the data directory, is `root/cifar100`.
"""
start = time.time()
task_path = assert_dirs(root, 'cifar100')
url = 'https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/cifar100/cifar-100-binary.tar.gz'
rq.files(url, gfile.path_join(task_path, url.split('/')[-1]))
with tarfile.open(gfile.path_join(task_path, url.split('/')[-1])) as t:
t.extractall(task_path)
noise_flie = gfile.listdir(task_path)
with open(gfile.path_join(task_path, 'train.bin'), 'rb') as fin:
data = np.frombuffer(fin.read(), dtype=np.uint8).reshape(-1, 3072 + 2)
train = data[:, 2:].reshape(-1, 3, 32, 32).transpose(0, 2, 3, 1)
train_label = data[:, 0 + fine_label].astype(np.int32)
for i in set(train_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for idx in range(train.shape[0]):
save_image(
gfile.path_join(task_path, 'train', str(train_label[idx]),
str(idx) + '.png'), array_to_image(train[idx]))
with open(gfile.path_join(task_path, 'test.bin'), 'rb') as fin:
data = np.frombuffer(fin.read(), dtype=np.uint8).reshape(-1, 3072 + 2)
test = data[:, 2:].reshape(-1, 3, 32, 32).transpose(0, 2, 3, 1)
test_label = data[:, 0 + fine_label].astype(np.int32)
for i in set(test_label):
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for idx in range(test.shape[0]):
save_image(
gfile.path_join(task_path, 'test', str(test_label[idx]),
str(idx) + '.png'), array_to_image(test[idx]))
for file in noise_flie:
gfile.remove(gfile.path_join(task_path, file))
print('cifar100 dataset download completed, run time %d min %.2f sec' %
divmod((time.time() - start), 60))
return task_path
def mnist(root):
"""MNIST handwritten digits dataset from http://yann.lecun.com/exdb/mnist
Each sample is an gray image (in 3D NDArray) with shape (28, 28, 1).
Attention: if exist dirs `root/mnist`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
mnist data:
`root/mnist/train/0/xx.png`
`root/mnist/train/2/xx.png`
`root/mnist/train/6/xx.png`
`root/mnist/test/0/xx.png`
`root/mnist/test/2/xx.png`
`root/mnist/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/mnist`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/mnist`.
"""
start = time.time()
task_path = assert_dirs(root, 'mnist')
url_list = ['https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/mnist/train-labels-idx1-ubyte.gz',
'https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/mnist/train-images-idx3-ubyte.gz',
'https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/mnist/t10k-labels-idx1-ubyte.gz',
'https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/mnist/t10k-images-idx3-ubyte.gz']
for url in url_list:
rq.files(url, gfile.path_join(task_path, url.split('/')[-1]))
with gzip.open(gfile.path_join(task_path, 'train-labels-idx1-ubyte.gz'), 'rb') as lbpath:
train_label = np.frombuffer(lbpath.read(), np.uint8, offset=8)
with gzip.open(gfile.path_join(task_path, 'train-images-idx3-ubyte.gz'), 'rb') as imgpath:
train = np.frombuffer(imgpath.read(), np.uint8, offset=16).reshape(len(train_label), 28, 28)
with gzip.open(gfile.path_join(task_path, 't10k-labels-idx1-ubyte.gz'), 'rb') as lbpath:
test_label = np.frombuffer(lbpath.read(), np.uint8, offset=8)
with gzip.open(gfile.path_join(task_path, 't10k-images-idx3-ubyte.gz'), 'rb') as imgpath:
test = np.frombuffer(imgpath.read(), np.uint8, offset=16).reshape(len(test_label), 28, 28)
for i in set(train_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for i in set(test_label):
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for idx in range(train.shape[0]):
save_image(gfile.path_join(task_path, 'train', str(train_label[idx]), str(idx)+'.png'),
array_to_image(train[idx].reshape(28, 28, 1)))
for idx in range(test.shape[0]):
save_image(gfile.path_join(task_path, 'test', str(test_label[idx]), str(idx)+'.png'),
array_to_image(test[idx].reshape(28, 28, 1)))
for url in url_list:
gfile.remove(gfile.path_join(task_path, url.split('/')[-1]))
print('mnist dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def mnist_kuzushiji10(root):
"""Kuzushiji-MNIST from https://github.com/rois-codh/kmnist.
Kuzushiji-MNIST is a drop-in replacement for the
MNIST dataset (28x28 grayscale, 70,000 images),
provided in the original MNIST format as well as a NumPy format.
Since MNIST restricts us to 10 classes, we chose one character to
represent each of the 10 rows of Hiragana when creating Kuzushiji-MNIST.
Each sample is an gray image (in 3D NDArray) with shape (28, 28, 1).
Attention: if exist dirs `root/mnist_kuzushiji10`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
mnist_kuzushiji10 data:
`root/mnist_kuzushiji10/train/0/xx.png`
`root/mnist_kuzushiji10/train/2/xx.png`
`root/mnist_kuzushiji10/train/6/xx.png`
`root/mnist_kuzushiji10/test/0/xx.png`
`root/mnist_kuzushiji10/test/2/xx.png`
`root/mnist_kuzushiji10/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/mnist_kuzushiji10`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/mnist_kuzushiji10`.
"""
start = time.time()
task_path = assert_dirs(root, 'mnist_kuzushiji10')
url_list = ['http://codh.rois.ac.jp/kmnist/dataset/kmnist/kmnist-train-imgs.npz',
'http://codh.rois.ac.jp/kmnist/dataset/kmnist/kmnist-train-labels.npz',
'http://codh.rois.ac.jp/kmnist/dataset/kmnist/kmnist-test-imgs.npz',
'http://codh.rois.ac.jp/kmnist/dataset/kmnist/kmnist-test-labels.npz']
for url in url_list:
rq.files(url, gfile.path_join(task_path, url.split('/')[-1]))
train = np.load(gfile.path_join(task_path, 'kmnist-train-imgs.npz'))['arr_0']
train_label = np.load(gfile.path_join(task_path, 'kmnist-train-labels.npz'))['arr_0']
test = np.load(gfile.path_join(task_path, 'kmnist-test-imgs.npz'))['arr_0']
test_label = np.load(gfile.path_join(task_path, 'kmnist-test-labels.npz'))['arr_0']
for i in set(train_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for i in set(test_label):
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for idx in range(train.shape[0]):
save_image(gfile.path_join(task_path, 'train', str(train_label[idx]), str(idx)+'.png'),
array_to_image(train[idx].reshape(28, 28, 1)))
for idx in range(test.shape[0]):
save_image(gfile.path_join(task_path, 'test', str(test_label[idx]), str(idx)+'.png'),
array_to_image(test[idx].reshape(28, 28, 1)))
for url in url_list:
gfile.remove(gfile.path_join(task_path, url.split('/')[-1]))
print('mnist_kuzushiji10 dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def mnist_kuzushiji49(root):
"""Kuzushiji-49 from https://github.com/rois-codh/kmnist.
Kuzushiji-49, as the name suggests, has 49 classes (28x28 grayscale, 270,912 images),
is a much larger, but imbalanced dataset containing 48 Hiragana
characters and one Hiragana iteration mark.
Each sample is an gray image (in 3D NDArray) with shape (28, 28, 1).
Attention: if exist dirs `root/mnist_kuzushiji49`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
mnist_kuzushiji49 data:
`root/mnist_kuzushiji49/train/0/xx.png`
`root/mnist_kuzushiji49/train/2/xx.png`
`root/mnist_kuzushiji49/train/6/xx.png`
`root/mnist_kuzushiji49/test/0/xx.png`
`root/mnist_kuzushiji49/test/2/xx.png`
`root/mnist_kuzushiji49/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/mnist_kuzushiji49`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/mnist_kuzushiji49`.
"""
start = time.time()
task_path = assert_dirs(root, 'mnist_kuzushiji49')
url_list = ['http://codh.rois.ac.jp/kmnist/dataset/k49/k49-train-imgs.npz',
'http://codh.rois.ac.jp/kmnist/dataset/k49/k49-train-labels.npz',
'http://codh.rois.ac.jp/kmnist/dataset/k49/k49-test-imgs.npz',
'http://codh.rois.ac.jp/kmnist/dataset/k49/k49-test-labels.npz']
for url in url_list:
rq.files(url, gfile.path_join(task_path, url.split('/')[-1]))
train = np.load(gfile.path_join(task_path, 'k49-train-imgs.npz'))['arr_0']
train_label = np.load(gfile.path_join(task_path, 'k49-train-labels.npz'))['arr_0']
test = np.load(gfile.path_join(task_path, 'k49-test-imgs.npz'))['arr_0']
test_label = np.load(gfile.path_join(task_path, 'k49-test-labels.npz'))['arr_0']
for i in set(train_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for i in set(test_label):
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for idx in range(train.shape[0]):
save_image(gfile.path_join(task_path, 'train', str(train_label[idx]), str(idx)+'.png'),
array_to_image(train[idx].reshape(28, 28, 1)))
for idx in range(test.shape[0]):
save_image(gfile.path_join(task_path, 'test', str(test_label[idx]), str(idx)+'.png'),
array_to_image(test[idx].reshape(28, 28, 1)))
for url in url_list:
gfile.remove(gfile.path_join(task_path, url.split('/')[-1]))
print('mnist_kuzushiji49 dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def mnist_tibetan(root):
"""Tibetan-MNIST from https://github.com/bat67/TibetanMNIST.
Tibetan-MNIST is a drop-in replacement for the
MNIST dataset (28x28 grayscale, 70,000 images),
provided in the original MNIST format as well as a NumPy format.
Since MNIST restricts us to 10 classes, we chose one character to
represent each of the 10 rows of Hiragana when creating Tibetan-MNIST.
Each sample is an gray image (in 3D NDArray) with shape (28, 28, 1).
Attention: if exist dirs `root/mnist_tibetan`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
mnist_tibetan data:
`root/mnist_tibetan/train/0/xx.png`
`root/mnist_tibetan/train/2/xx.png`
`root/mnist_tibetan/train/6/xx.png`
`root/mnist_tibetan/test/0/xx.png`
`root/mnist_tibetan/test/2/xx.png`
`root/mnist_tibetan/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/mnist_tibetan`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/mnist_tibetan`.
"""
start = time.time()
print('Downloading data from https://github.com/Hourout/datasets/tree/master/TibetanMNIST')
task_path = assert_dirs(root, 'mnist_tibetan')
url_list = ['https://raw.githubusercontent.com/Hourout/datasets/master/TibetanMNIST/TibetanMNIST_28_28_01.csv',
'https://raw.githubusercontent.com/Hourout/datasets/master/TibetanMNIST/TibetanMNIST_28_28_02.csv']
data = pd.DataFrame()
for url in url_list:
s = requests.get(url).content
data = pd.concat([data, pd.read_csv(io.StringIO(s.decode('utf-8')),header=None, dtype='uint8')])
train = data.loc[:, 1:].values.reshape(-1, 28, 28)
train_label = data.loc[:, 0].values
for i in set(train_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for idx in range(train.shape[0]):
save_image(gfile.path_join(task_path, 'train', str(train_label[idx]), str(idx)+'.png'),
array_to_image(train[idx].reshape(28, 28, 1)))
print('mnist_tibetan dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def mnist_kannada(root):
"""kannada-MNIST from https://github.com/vinayprabhu/Kannada_MNIST.
The Kannada-MNIST dataset was created an a drop-in substitute for the standard MNIST dataset.
Each sample is an gray image (in 3D NDArray) with shape (28, 28, 1).
Attention: if exist dirs `root/mnist_kannada`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
mnist_kannada data:
`root/mnist_kannada/train/0/xx.png`
`root/mnist_kannada/train/2/xx.png`
`root/mnist_kannada/train/6/xx.png`
`root/mnist_kannada/test/0/xx.png`
`root/mnist_kannada/test/2/xx.png`
`root/mnist_kannada/test/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/mnist_kannada`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/mnist_kannada`.
"""
start = time.time()
print('Downloading data from https://github.com/Hourout/datasets/releases/download/0.0.1/kannada_MNIST.zip')
task_path = assert_dirs(root, 'mnist_kannada')
zip_path = rq.files('https://github.com/Hourout/datasets/releases/download/0.0.1/kannada_MNIST.zip', task_path+'/kannada_MNIST.zip')
unzip_path = un_zip(task_path+'/kannada_MNIST.zip')
train = pd.read_csv(gfile.path_join(task_path, 'kannada_MNIST/kannada_MNIST_train.csv'), header=None, dtype='uint8')
test = pd.read_csv(gfile.path_join(task_path, 'kannada_MNIST/kannada_MNIST_test.csv'), header=None, dtype='uint8')
for i in set(train[0]):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for i in range(len(train)):
save_image(gfile.path_join(task_path, 'train', str(train.iat[i, 0]), str(i)+'.png'),
array_to_image(train.iloc[i, 1:].values.reshape(28, 28, 1)))
for i in range(len(test)):
save_image(gfile.path_join(task_path, 'test', str(test.iat[i, 0]), str(i)+'.png'),
array_to_image(test.iloc[i, 1:].values.reshape(28, 28, 1)))
gfile.remove(zip_path)
gfile.remove(unzip_path)
print('mnist_kannada dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def coil100(root):
"""COIL100 dataset from http://www.cs.columbia.edu/CAVE/software/softlib/coil-100.php
"Columbia Object Image Library (COIL-100),"
S. A. Nene, S. K. Nayar and H. Murase,
Technical Report CUCS-006-96, February 1996.
Each sample is an gray image (in 3D NDArray) with shape (128, 128, 1).
Attention: if exist dirs `root/coil100`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
coil100 data:
`root/coil100/train/0/xx.png`
`root/coil100/train/2/xx.png`
`root/coil100/train/6/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/coil100`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/coil100`.
"""
start = time.time()
task_path = assert_dirs(root, 'coil100')
url = "http://www.cs.columbia.edu/CAVE/databases/SLAM_coil-20_coil-100/coil-100/coil-100.zip"
rq.files(url, gfile.path_join(task_path, 'coil100.zip'))
un_zip(gfile.path_join(task_path, 'coil100.zip'))
image = gfile.listdir(gfile.path_join(task_path, 'coil100', 'coil-100'))
t = pd.DataFrame(image, columns=['image'])
t['label'] = t.image.map(lambda x:x.split('__')[0][3:])
t['image_old_path'] = t.image.map(lambda x:gfile.path_join(task_path, 'coil100', 'coil-100', x))
t['image_new_path'] = (t.label+'/'+t.image).map(lambda x:gfile.path_join(task_path, 'train', x))
for i in t.label.unique():
gfile.makedirs(gfile.path_join(task_path, 'train', i))
for i,j in zip(t.image_old_path, t.image_new_path):
gfile.copy(i, j)
gfile.remove(gfile.path_join(task_path, 'coil100.zip'))
gfile.remove(gfile.path_join(task_path, 'coil100'))
gfile.remove(gfile.path_join(task_path, 'train', 'vertGroupppm2png.pl'))
print('coil100 dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path
def assert_dirs(root, root_dir=None, delete=True, make_root_dir=True):
assert gfile.isdir(root), '{} should be directory.'.format(root)
if root_dir is not None:
assert isinstance(root_dir, str), '{} should be str.'.format(root_dir)
task_path = gfile.path_join(root, root_dir)
if gfile.exists(task_path):
if delete:
gfile.remove(task_path)
gfile.makedirs(task_path)
else:
if make_root_dir:
gfile.makedirs(task_path)
return task_path
else:
if not gfile.exists(root):
gfile.makedirs(root)
return root
def stl10(root):
"""Stl10 dataset from http://ai.stanford.edu/~acoates/stl10
The STL-10 dataset is an image recognition dataset for developing
unsupervised feature learning, deep learning, self-taught learning algorithms.
It is inspired by the CIFAR-10 dataset but with some modifications.
In particular, each class has fewer labeled training examples than in CIFAR-10,
but a very large set of unlabeled examples is provided to learn image models
prior to supervised training. The primary challenge is to make use of the
unlabeled data (which comes from a similar but different
distribution from the labeled data) to build a useful prior.
We also expect that the higher resolution of this dataset (96x96)
will make it a challenging benchmark for developing
more scalable unsupervised learning methods.
Attention: if exist dirs `root/stl10`, api will delete it and create it.
Data storage directory:
root = `/user/.../mydata`
stl10 data:
`root/stl10/train/1/xx.png`
`root/stl10/train/4/xx.png`
`root/stl10/train/8/xx.png`
`root/stl10/test/1/xx.png`
`root/stl10/test/4/xx.png`
`root/stl10/test/8/xx.png`
Args:
root: str, Store the absolute path of the data directory.
example:if you want data path is `/user/.../mydata/stl10`,
root should be `/user/.../mydata`.
Returns:
Store the absolute path of the data directory, is `root/stl10`.
"""
start = time.time()
task_path = assert_dirs(root, 'stl10')
url = "http://ai.stanford.edu/~acoates/stl10/stl10_binary.tar.gz"
rq.files(url, gfile.path_join(task_path, url.split('/')[-1]))
un_tar(un_gz(gfile.path_join(task_path, url.split('/')[-1])))
with open(gfile.path_join(task_path, 'stl10_binary/stl10_binary/test_X.bin'), 'rb') as fin:
data = np.frombuffer(fin.read(), dtype=np.uint8).reshape(-1, 3,96,96).transpose((0, 3, 2, 1))
with open(gfile.path_join(task_path, 'stl10_binary/stl10_binary/test_y.bin'), 'rb') as fin:
data_label = np.frombuffer(fin.read(), dtype=np.uint8)
for i in set(data_label):
gfile.makedirs(gfile.path_join(task_path, 'test', str(i)))
for idx in range(data.shape[0]):
save_image(gfile.path_join(task_path, 'test', str(data_label[idx]), str(idx)+'.png'), array_to_image(data[idx]))
with open(gfile.path_join(task_path, 'stl10_binary/stl10_binary/train_X.bin'), 'rb') as fin:
data = np.frombuffer(fin.read(), dtype=np.uint8).reshape(-1, 3,96,96).transpose((0, 3, 2, 1))
with open(gfile.path_join(task_path, 'stl10_binary/stl10_binary/train_y.bin'), 'rb') as fin:
data_label = np.frombuffer(fin.read(), dtype=np.uint8)
for i in set(data_label):
gfile.makedirs(gfile.path_join(task_path, 'train', str(i)))
for idx in range(data.shape[0]):
save_image(gfile.path_join(task_path, 'train', str(data_label[idx]), str(idx)+'.png'), array_to_image(data[idx]))
with open(gfile.path_join(task_path, 'stl10_binary/stl10_binary/unlabeled_X.bin'), 'rb') as fin:
data = np.frombuffer(fin.read(), dtype=np.uint8).reshape(-1, 3,96,96).transpose((0, 3, 2, 1))
gfile.makedirs(gfile.path_join(task_path, 'unlabeled'))
for idx in range(data.shape[0]):
save_image(gfile.path_join(task_path, 'unlabeled', str(idx)+'.png'), array_to_image(data[idx]))
gfile.remove(gfile.path_join(task_path, 'stl10_binary.tar.gz'))
gfile.remove(gfile.path_join(task_path, 'stl10_binary.tar'))
gfile.remove(path_join(task_path, 'stl10_binary'))
print('stl10 dataset download completed, run time %d min %.2f sec' %divmod((time.time()-start), 60))
return task_path