def load_syn(scale=True, usps=False, all_use=False):
syn_data = loadmat(base_dir + '/syn_number.mat')
syn_train = syn_data['train_data']
syn_test = syn_data['test_data']
syn_train = syn_train.transpose(0, 3, 1, 2).astype(np.float32)
syn_test = syn_test.transpose(0, 3, 1, 2).astype(np.float32)
syn_labels_train = syn_data['train_label']
syn_labels_test = syn_data['test_label']
train_label = syn_labels_train
inds = np.random.permutation(syn_train.shape[0])
syn_train = syn_train[inds]
train_label = train_label[inds]
test_label = syn_labels_test
# syn_train = syn_train[:25000]
# train_label = train_label[:25000]
# syn_test = syn_test[:9000]
# test_label = test_label[:9000]
train_label = dense_to_one_hot(train_label)
test_label = dense_to_one_hot(test_label)
print('syn number train X shape->', syn_train.shape)
print('syn number train y shape->', train_label.shape)
print('syn number test X shape->', syn_test.shape)
print('syn number test y shape->', test_label.shape)
return syn_train, train_label, syn_test, test_label
def load_syn(base_dir, scale=32):
syn_train = loadmat(base_dir +
'/synth_train_{}x{}.mat'.format(scale, scale))
syn_test = loadmat(base_dir + '/synth_test_{}x{}.mat'.format(scale, scale))
# train
syn_train_im = syn_train['X']
syn_train_im = syn_train_im.transpose(3, 2, 0, 1).astype(np.float32)
train_label = dense_to_one_hot(syn_train['y'].squeeze())
# idx = np.random.permutation(syn_train_im.shape[0])
# syn_train_im = syn_train_im[idx]
# train_label = train_label[idx]
# syn_train_im = syn_train_im[:25000]
# train_label = train_label[:25000]
# test
syn_test_im = syn_test['X']
syn_test_im = syn_test_im.transpose(3, 2, 0, 1).astype(np.float32)
test_label = dense_to_one_hot(syn_test['y'].squeeze())
# syn_test_im = syn_test_im[:9000]
# test_label = test_label[:9000]
print('syn number train X shape->', syn_train_im.shape)
print('syn number train y shape->', train_label.shape)
print('syn number test X shape->', syn_test_im.shape)
print('syn number test y shape->', test_label.shape)
print("====================")
return syn_train_im, train_label, syn_test_im, test_label
def load_usps(base_dir):
dataset = loadmat(base_dir + '/usps_28x28.mat')
data_set = dataset['dataset']
img_train = data_set[0][0]
label_train = data_set[0][1]
img_test = data_set[1][0]
label_test = data_set[1][1]
inds = np.random.permutation(img_train.shape[0])
img_train = img_train[inds]
label_train = label_train[inds]
img_train = img_train * 255
img_test = img_test * 255
img_train = img_train.reshape((img_train.shape[0], 1, 28, 28))
img_test = img_test.reshape((img_test.shape[0], 1, 28, 28))
label_train = dense_to_one_hot(label_train)
label_test = dense_to_one_hot(label_test)
img_train = np.concatenate([img_train, img_train, img_train], 1)
img_test = np.concatenate([img_test, img_test, img_test], 1)
print('usps train X shape->', img_train.shape)
print('usps train y shape->', label_train.shape)
print('usps test X shape->', img_test.shape)
print('usps test y shape->', label_test.shape)
print("====================")
return img_train, label_train, img_test, label_test
def load_svhn(base_dir, scale=32):
svhn_train = loadmat(base_dir + '/svhn_train_{}x{}.mat'.format(scale, scale))
svhn_test = loadmat(base_dir + '/svhn_test_{}x{}.mat'.format(scale, scale))
svhn_train_im = svhn_train['X']
svhn_train_im = svhn_train_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label = dense_to_one_hot(svhn_train['y'])
# sample train set
idx = np.random.permutation(svhn_train_im.shape[0])
svhn_train_im = svhn_train_im[idx]
svhn_label = svhn_label[idx]
svhn_train_im = svhn_train_im[:25000]
svhn_label = svhn_label[:25000]
svhn_test_im = svhn_test['X']
svhn_test_im = svhn_test_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label_test = dense_to_one_hot(svhn_test['y'])
# sample test set
svhn_test_im = svhn_test_im[:14459]
svhn_label_test = svhn_label_test[:14459]
print('svhn train X shape->', svhn_train_im.shape)
print('svhn train y shape->', svhn_label.shape)
print('svhn test X shape->', svhn_test_im.shape)
print('svhn test y shape->', svhn_label_test.shape)
print("====================")
return svhn_train_im, svhn_label, svhn_test_im, svhn_label_test
def load_usps(directory,all_use=False):
base_dir = directory
#f = gzip.open('data/usps_28x28.pkl', 'rb')
#data_set = pickle.load(f)
#f.close()
dataset = loadmat(base_dir + '/usps_28x28.mat')
data_set = dataset['dataset']
img_train = data_set[0][0]
label_train = data_set[0][1]
img_test = data_set[1][0]
label_test = data_set[1][1]
inds = np.random.permutation(img_train.shape[0])
img_train = img_train[inds]
label_train = label_train[inds]
img_train = img_train * 255
img_test = img_test * 255
img_train = img_train.reshape((img_train.shape[0], 1, 28, 28))
img_test = img_test.reshape((img_test.shape[0], 1, 28, 28))
#img_test = dense_to_one_hot(img_test)
label_train = dense_to_one_hot(label_train)
label_test = dense_to_one_hot(label_test)
img_train = np.concatenate([img_train, img_train, img_train, img_train], 0)
label_train = np.concatenate([label_train, label_train, label_train, label_train], 0)
print('usps train X shape->', img_train.shape)
print('usps train y shape->', label_train.shape)
print('usps test X shape->', img_test.shape)
print('usps test y shape->', label_test.shape)
return img_train, label_train, img_test, label_test
def load_svhn():
svhn_train = loadmat('../data/train_32x32.mat')
svhn_test = loadmat('../data/test_32x32.mat')
svhn_train_im = svhn_train['X']
svhn_train_im = svhn_train_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label = dense_to_one_hot(svhn_train['y'])
svhn_test_im = svhn_test['X']
svhn_test_im = svhn_test_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label_test = dense_to_one_hot(svhn_test['y'])
return svhn_train_im, svhn_label, svhn_test_im, svhn_label_test
def load_syn(scale=True, usps=False, all_use=False):
syn_train = loadmat(base_dir + '/synth_train_28x28.mat')
syn_test = loadmat(base_dir + '/synth_test_28x28.mat')
syn_train_im = syn_train['X']
syn_train_im = syn_train_im.transpose(3, 2, 0, 1).astype(np.float32)
train_label = dense_to_one_hot(syn_train['y'])
syn_test_im = syn_test['X']
syn_test_im = syn_test_im.transpose(3, 2, 0, 1).astype(np.float32)
test_label = dense_to_one_hot(syn_test['y'])
print('syn number train X shape->', syn_train_im.shape)
print('syn number train y shape->', train_label.shape)
print('syn number test X shape->', syn_test_im.shape)
print('syn number test y shape->', test_label.shape)
return syn_train_im, train_label, syn_test_im, test_label
def load_svhn():
svhn_train = loadmat(base_dir + '/svhn_train_28x28.mat')
svhn_test = loadmat(base_dir + '/svhn_test_28x28.mat')
svhn_train_im = svhn_train['X']
svhn_train_im = svhn_train_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label = dense_to_one_hot(svhn_train['y'])
svhn_test_im = svhn_test['X']
svhn_test_im = svhn_test_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label_test = dense_to_one_hot(svhn_test['y'])
print('svhn train X shape->', svhn_train_im.shape)
print('svhn train y shape->', svhn_label.shape)
print('svhn test X shape->', svhn_test_im.shape)
print('svhn test y shape->', svhn_label_test.shape)
return svhn_train_im, svhn_label, svhn_test_im, svhn_label_test
def load_svhn():
svhn_train = loadmat(base_dir + '/train_32x32.mat')
svhn_test = loadmat(base_dir + '/test_32x32.mat')
svhn_train_im = svhn_train['X']
svhn_train_im = svhn_train_im.transpose(3, 2, 0, 1).astype(np.float32)
print('svhn train y shape before dense_to_one_hot->',
svhn_train['y'].shape)
svhn_label = dense_to_one_hot(svhn_train['y'])
print('svhn train y shape after dense_to_one_hot->', svhn_label.shape)
svhn_test_im = svhn_test['X']
svhn_test_im = svhn_test_im.transpose(3, 2, 0, 1).astype(np.float32)
svhn_label_test = dense_to_one_hot(svhn_test['y'])
svhn_train_im = svhn_train_im[:25000]
svhn_label = svhn_label[:25000]
svhn_test_im = svhn_test_im[:9000]
svhn_label_test = svhn_label_test[:9000]
print('svhn train X shape->', svhn_train_im.shape)
print('svhn train y shape->', svhn_label.shape)
print('svhn test X shape->', svhn_test_im.shape)
print('svhn test y shape->', svhn_label_test.shape)
return svhn_train_im, svhn_label, svhn_test_im, svhn_label_test
