def test_pca_input(sh, m):
"""
Generate input data from labeled faces data
:param sh: The shape of the image tiles to sample
:param m: The number of examples to sample
:return: 2d ndarray where columns are examples
"""
return sample_patches(load_lfwc(), patch_shape=sh, n_samples=m, flatten=True).T
def get_input(input_pattern=lfwc_pattern,
n_samples=5000,
patch_shape=(12, 12),
n_images=-1,
epsilon=0.1,
norm_axis=0):
"""
Get input data from lfwc patches
:param input_pattern: The pattern to glob for image files
:param n_images: The number of images to draw samples from
:param n_samples: The number of samples to extract
:param patch_shape: The shape of the patches to sample
:param epsilon: Regularization for whitening, or None to disable whitening
:param norm_axis: Whether to mean normalize across each feature (0) or each patch (1)
:return: m x n ndarray of flattened images
"""
paths = glob.glob(input_pattern)
random.shuffle(paths)
paths = paths[:n_images]
print("Sampling {} {}x{} patches from {} images...".format(
n_samples, patch_shape[0], patch_shape[1], len(paths)))
# read images into a m x img_row x img_col ndarray
imgs = np.asarray([cv2.imread(path, 0) for path in paths],
dtype=theano.config.floatX)
imgs /= 255. # scale to [0,1]
if input_pattern == lfwc_pattern and patch_shape == (
64, 64) and n_samples == lfwc_count:
patches = load_matrix_2d(paths)
epsilon = 0 # disable whitening... too slow
else:
# sample patches into a 3d array
patches = sample_patches(imgs,
patch_shape=patch_shape,
n_samples=n_samples)
# flatten each image to 1d
patches = patches.reshape(patches.shape[0],
patch_shape[0] * patch_shape[1])
if norm_axis == 0:
# subtract per feature (pixel) means across the data set (m x n)
normed = patches - patches.mean(axis=0)
else:
# remove mean val of each patch
normed = patches - patches.mean(axis=1)[:, np.newaxis]
if epsilon is None:
return normed
return whiten_zca(normed, epsilon=epsilon)
def test_pca_input(sh, m):
"""
Generate input data from labeled faces data
:param sh: The shape of the image tiles to sample
:param m: The number of examples to sample
:return: 2d ndarray where columns are examples
"""
return sample_patches(load_lfwc(),
patch_shape=sh,
n_samples=m,
flatten=True).T
def main(cmd, args):
"""process the command"""
data_set = "lfw" if len(args) > 0 and args[0] == "lfw" else "lfwc"
if cmd == "show":
if len(args) > 0 and args[0] == "lfw":
show_images(lfw_paths())
else:
show_images(lfwc_paths())
return
n_samples = 10000
sh = (20, 20)
out_shape = (8, 8)
data = load_lfw() if data_set == "lfw" else load_lfwc()
out_dir = "./img/"
shape_range = range(8, 21, 2)
if cmd == "sample":
patches = sample_patches(data, patch_shape=sh, n_samples=n_samples)
tile_images(patches, patch_shape=sh, output_shape=out_shape, show=True)
elif cmd == "sweep":
shapes = [(i, i) for i in shape_range]
for sh in shapes:
patches = sample_patches(data, n_samples=n_samples, patch_shape=sh)
img = tile_images(patches, sh, output_shape=out_shape, show=True)
fn = os.path.join(out_dir,
"{}_patches_{}.png".format(data_set, sh))
print("saving {}...".format(fn))
img.save(fn)
elif cmd == "alfw":
print(load_alfw())
else:
print("unexpected cmd")
def main(cmd, args):
"""process the command"""
data_set = "lfw" if len(args) > 0 and args[0] == "lfw" else "lfwc"
if cmd == "show":
if len(args) > 0 and args[0] == "lfw":
show_images(lfw_paths())
else:
show_images(lfwc_paths())
return
n_samples = 10000
sh = (20, 20)
out_shape = (8, 8)
data = load_lfw() if data_set == "lfw" else load_lfwc()
out_dir = "./img/"
shape_range = range(8, 21, 2)
if cmd == "sample":
patches = sample_patches(data, patch_shape=sh, n_samples=n_samples)
tile_images(patches, patch_shape=sh, output_shape=out_shape, show=True)
elif cmd == "sweep":
shapes = [(i, i) for i in shape_range]
for sh in shapes:
patches = sample_patches(data, n_samples=n_samples, patch_shape=sh)
img = tile_images(patches, sh, output_shape=out_shape, show=True)
fn = os.path.join(out_dir, "{}_patches_{}.png".format(data_set, sh))
print("saving {}...".format(fn))
img.save(fn)
elif cmd == "alfw":
print(load_alfw())
else:
print("unexpected cmd")
def get_input(input_pattern=lfwc_pattern,
n_samples=5000, patch_shape=(12, 12),
n_images=-1, epsilon=0.1, norm_axis=0):
"""
Get input data from lfwc patches
:param input_pattern: The pattern to glob for image files
:param n_images: The number of images to draw samples from
:param n_samples: The number of samples to extract
:param patch_shape: The shape of the patches to sample
:param epsilon: Regularization for whitening, or None to disable whitening
:param norm_axis: Whether to mean normalize across each feature (0) or each patch (1)
:return: m x n ndarray of flattened images
"""
paths = glob.glob(input_pattern)
random.shuffle(paths)
paths = paths[:n_images]
print("Sampling {} {}x{} patches from {} images..."
.format(n_samples, patch_shape[0], patch_shape[1], len(paths)))
# read images into a m x img_row x img_col ndarray
imgs = np.asarray([cv2.imread(path, 0) for path in paths], dtype=theano.config.floatX)
imgs /= 255. # scale to [0,1]
if input_pattern == lfwc_pattern and patch_shape == (64,64) and n_samples == lfwc_count:
patches = load_matrix_2d(paths)
epsilon = 0 # disable whitening... too slow
else:
# sample patches into a 3d array
patches = sample_patches(imgs, patch_shape=patch_shape, n_samples=n_samples)
# flatten each image to 1d
patches = patches.reshape(patches.shape[0], patch_shape[0]*patch_shape[1])
if norm_axis == 0:
# subtract per feature (pixel) means across the data set (m x n)
normed = patches - patches.mean(axis=0)
else:
# remove mean val of each patch
normed = patches - patches.mean(axis=1)[:, np.newaxis]
if epsilon is None:
return normed
return whiten_zca(normed, epsilon=epsilon)