def ComputeFeatures(self):
"""Compute SVM feature vectors for all images.
Generally, you do not need to call this method yourself, as it will be
called automatically by :meth:`RunSvm`.
"""
if self.train_images == None or self.test_images == None:
sys.exit("Please specify the corpus.")
train_sizes = map(len, self.train_images)
train_size = sum(train_sizes)
test_sizes = map(len, self.test_images)
test_size = sum(test_sizes)
train_images = util.UngroupLists(self.train_images)
test_images = util.UngroupLists(self.test_images)
images = train_images + test_images
start_time = time.time()
# Compute features for all images.
features = self.GetImageFeatures(images, block=True)
self.compute_feature_time = time.time() - start_time
# Split results by training/testing set
train_features, test_features = util.SplitList(features,
[train_size, test_size])
# Split training set by class
train_features = util.SplitList(train_features, train_sizes)
# Split testing set by class
test_features = util.SplitList(test_features, test_sizes)
# Store features as list of 2D arrays
self.train_features = [
np.array(f, util.ACTIVATION_DTYPE) for f in train_features
]
self.test_features = [
np.array(f, util.ACTIVATION_DTYPE) for f in test_features
]
def ImprintS2Prototypes(self, num_prototypes):
"""Imprint a set of S2 prototypes from the training images.
Patches are drawn from all classes of the training data.
:param int num_prototypes: The number of C1 patches to sample.
"""
if self.train_images == None:
sys.exit("Please specify the training corpus before imprinting "
"prototypes.")
model = self.model
start_time = time.time()
image_files = util.UngroupLists(self.train_images)
# Represent each image file as an empty model state.
input_states = map(model.MakeStateFromFilename, image_files)
try:
prototypes, locations = misc.ImprintKernels(
model,
model.LayerClass.C1,
model.s2_kernel_sizes,
num_prototypes,
input_states,
normalize=model.s2_kernels_are_normed,
pool=self.pool)
except InputSourceLoadException, ex:
if ex.source != None:
path = ex.source.image_path
else:
path = '<unknown>'
logging.error("Failed to process image (%s): image read error",
path)
sys.exit(-1)
def PrepareFeatures(features_per_class):
"""Format feature vectors for use by the sklearn library.
:param features_per_class: Per-class feature vectors.
:type features_per_class: 2D array-like
:returns: Feature vectors and labels.
:rtype: 2D-ndarray and 1D-ndarray
"""
num_classes = len(features_per_class)
if num_classes == 2:
labels = [1, -1]
else:
labels = range(1, num_classes + 1)
class_sizes = map(len, features_per_class)
labels_per_class = [[label] * size
for label, size in zip(labels, class_sizes)]
features = util.UngroupLists(features_per_class)
features = np.array(features)
labels = util.UngroupLists(labels_per_class)
labels = np.array(labels)
return features, labels