def __init__(self, batch_size=1000, train=True):
self.batch_size = batch_size
self.classes = 10
if train:
self.images, self.labels, _, _ = mnist_web.mnist(path='.')
else:
_, _, self.images, self.labels = mnist_web.mnist(path='.')
self.n = self.images.shape[0]
self.images *= 255
self.images = self.images.astype(int)
self.images[self.images <= 128] = -1
self.images[self.images > 128] = 1
def __init__(self, train=True, cuda=False):
if train:
self.images, self.labels, _, _ = mnist_web.mnist(path='.')
else:
_, _, self.images, self.labels = mnist_web.mnist(path='.')
self.images *= 255
self.images = self.images.astype('int32')
#self.labels = np.sum(self.labels * np.arange(0,10),1).reshape(-1,1)
self.labels = self.labels.astype('int32')
self.images[self.images <= 128] = -1
self.images[self.images > 128] = 1
self.images = torch.from_numpy(self.images)
self.labels = torch.from_numpy(self.labels)
if cuda:
self.images = self.images.float().cuda()
self.labels = self.labels.float().cuda()
def __init__(self, train = True, margin = 0, noise_rate = 0):
if train:
self.train = True
self.images, self.labels, _, _= mnist_web.mnist(path='.')
self.MARGIN = margin
self.noise_rate = noise_rate
else:
self.train = False
self.MARGIN = 0
self.noise_rate = 0
_, _, self.images, self.labels = mnist_web.mnist(path='.')
#self.labels = np.sum(self.labels * np.arange(0,10),1).reshape(-1,1)
self.len = self.labels.shape[0]
self.images *= 255
self.images[self.images<=128] = -1
self.images[self.images>128] = 1
self.images = torch.from_numpy(self.images).type(DTYPE)
self.images_raw= self.images.clone()
self.labels = torch.from_numpy(self.labels).type(DTYPE)
self.images = self.images.reshape(-1, IMAGE_R, IMAGE_R)
self.images = torch.nn.functional.pad(self.images, \
(self.MARGIN,self.MARGIN,self.MARGIN,self.MARGIN), "constant", -1)
def load_data_wrapper2(path):
train_images, train_labels, test_images, test_labels = mnist_web.mnist(
path)
training_inputs = [np.reshape(x, (784, 1)) for x in train_images]
training_results = [vectorized_result2(y) for y in train_labels]
training_data = list(zip(training_inputs, training_results))
# validation_inputs = [np.reshape(x, (784, 1)) for x in test_images]
# validation_data = zip(validation_inputs, va_d[1])
test_inputs = [
np.reshape(x, (784, 1)) for x in test_images
] # np.concatenate([np.reshape(x, (784, 1)) for x in test_images], axis=1)
test_results = [vectorized_result3(y) for y in test_labels]
test_data = list(zip(test_inputs, test_results))
return training_data, test_data