def generate_synthetic(self):
"""
Generate a synthetic dataset of 4 classes that follows some simple
but strong cooccurence rules.
"""
num_images = 1000
choice_probs = {
(1, 0, 0, 0): 2,
(0, 1, 0, 0): 1,
(0, 0, 1, 0): 1,
(0, 0, 0, 1): 1,
(1, 1, 0, 0): 2,
(1, 0, 1, 0): 0,
(1, 0, 0, 1): 0,
(0, 1, 1, 0): 0,
(0, 1, 0, 1): 1,
(0, 0, 1, 1): 2,
(1, 1, 1, 0): 0,
(1, 1, 0, 1): 1,
(1, 0, 1, 1): 1,
(0, 1, 1, 1): 2,
}
probs = np.array(choice_probs.values())
cum_probs = np.cumsum(1.0 * probs / np.sum(probs))
self.classes = ["A", "B", "C", "D"]
for i in range(0, num_images):
image = Image(100, 100, self.classes, str(i))
choice = np.where(cum_probs > np.random.rand())[0][0]
objects = []
for cls_ind, clas in enumerate(choice_probs.keys()[choice]):
if clas == 1:
objects.append(np.array([0, 0, 0, 0, cls_ind, 0, 0]))
image.objects_table = Table(np.array(objects), Image.columns)
self.images.append(image)
def generate_synthetic(self):
"""
Generate a synthetic dataset of 4 classes that follows some simple
but strong cooccurence rules.
"""
num_images = 1000
choice_probs = {
(1, 0, 0, 0): 2,
(0, 1, 0, 0): 1,
(0, 0, 1, 0): 1,
(0, 0, 0, 1): 1,
(1, 1, 0, 0): 2,
(1, 0, 1, 0): 0,
(1, 0, 0, 1): 0,
(0, 1, 1, 0): 0,
(0, 1, 0, 1): 1,
(0, 0, 1, 1): 2,
(1, 1, 1, 0): 0,
(1, 1, 0, 1): 1,
(1, 0, 1, 1): 1,
(0, 1, 1, 1): 2
}
probs = np.array(choice_probs.values())
cum_probs = np.cumsum(1. * probs / np.sum(probs))
self.classes = ['A', 'B', 'C', 'D']
for i in range(0, num_images):
image = Image(100, 100, self.classes, str(i))
choice = np.where(cum_probs > np.random.rand())[0][0]
objects = []
for cls_ind, clas in enumerate(choice_probs.keys()[choice]):
if clas == 1:
objects.append(np.array([0, 0, 0, 0, cls_ind, 0, 0]))
image.objects_table = Table(np.array(objects), Image.columns)
self.images.append(image)
def test_expected_time(self): d = PerfectDetector(self.train_dataset,self.train_dataset,'dog') img = Image(500,375,self.train_dataset.classes,'test') print d.expected_time(img) assert(d.expected_time(img) == 10) img = Image(250,375,self.train_dataset.classes,'test') assert(d.expected_time(img) == 5)
def test_get_random_windows(self):
image = Image(width=3,height=2,classes=[],name='test')
window_params = WindowParams(
min_width=2,stride=1,scales=[1,0.5],aspect_ratios=[0.5])
windows = image.get_random_windows(window_params,2)
assert(windows.shape[0] == 2)
windows = image.get_random_windows(window_params,3)
assert(windows.shape[0] == 3)
def test_get_random_windows(self):
image = Image(width=3, height=2, classes=[], name='test')
window_params = WindowParams(min_width=2,
stride=1,
scales=[1, 0.5],
aspect_ratios=[0.5])
windows = image.get_random_windows(window_params, 2)
assert (windows.shape[0] == 2)
windows = image.get_random_windows(window_params, 3)
assert (windows.shape[0] == 3)
def test_get_windows_lots(self):
t = time.time()
image = Image(width=640,height=480,classes=[],name='test')
window_params = WindowParams()
window_params.min_width=10
window_params.stride=8
window_params.aspect_ratios=[0.5,1,1.5]
window_params.scales=1./2**np.array([0,0.5,1,1.5,2])
print(window_params)
windows = image.get_windows(window_params)
time_passed = time.time()-t
print("Generating windows took %.3f seconds"%time_passed)
print(np.shape(windows))
print(windows[:10,:])
rand_ind = np.random.permutation(np.shape(windows)[0])[:10]
print(windows[rand_ind,:])
def load_from_json(self, filename):
"Load all parameters of the dataset from a JSON file."
with open(filename) as f:
config = json.load(f)
self.classes = config['classes']
for data in config['images']:
self.images.append(Image.load_from_json_data(self.classes, data))
def test_get_windows_lots(self):
t = time.time()
image = Image(width=640, height=480, classes=[], name='test')
window_params = WindowParams()
window_params.min_width = 10
window_params.stride = 8
window_params.aspect_ratios = [0.5, 1, 1.5]
window_params.scales = 1. / 2**np.array([0, 0.5, 1, 1.5, 2])
print(window_params)
windows = image.get_windows(window_params)
time_passed = time.time() - t
print("Generating windows took %.3f seconds" % time_passed)
print(np.shape(windows))
print(windows[:10, :])
rand_ind = np.random.permutation(np.shape(windows)[0])[:10]
print(windows[rand_ind, :])
def load_from_json(self, filename):
"Load all parameters of the dataset from a JSON file."
with open(filename) as f:
config = json.load(f)
self.classes = config["classes"]
for data in config["images"]:
self.images.append(Image.load_from_json_data(self.classes, data))
def test_get_det_gt(self):
image = Image.load_from_json_data(self.classes, self.data)
objects_table = Table(
np.array([[0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0],
[2, 2, 2, 2, 2, 0, 0]]), self.columns)
assert (image.get_objects() == objects_table)
data = self.data.copy()
data['objects'][0]['diff'] = 1
data['objects'][1]['trun'] = 1
image = Image.load_from_json_data(self.classes, data)
objects_table = Table(
np.array([[0, 0, 0, 0, 0, 1, 0], [1, 1, 1, 1, 1, 0, 1],
[2, 2, 2, 2, 2, 0, 0]]), self.columns)
assert (image.get_objects(with_diff=True,
with_trun=True) == objects_table)
objects_table = Table(
np.array([[1, 1, 1, 1, 1, 0, 1], [2, 2, 2, 2, 2, 0, 0]]),
self.columns)
assert (image.get_objects(with_diff=False,
with_trun=True) == objects_table)
# this should be default behavior
assert (image.get_objects() == objects_table)
objects_table = Table(np.array([[2, 2, 2, 2, 2, 0, 0]]), self.columns)
assert (image.get_objects(with_diff=False,
with_trun=False) == objects_table)
objects_table = Table(
np.array([[0, 0, 0, 0, 0, 1, 0], [2, 2, 2, 2, 2, 0, 0]]),
self.columns)
assert (image.get_objects(with_diff=True,
with_trun=False) == objects_table)
# What if everything is filtered out?
data['objects'] = data['objects'][:-1]
objects_table = Table(
np.array([[0, 0, 0, 0, 0, 1, 0], [1, 1, 1, 1, 1, 0, 1]]),
self.columns)
image = Image.load_from_json_data(self.classes, data)
assert (image.get_objects(with_diff=True,
with_trun=True) == objects_table)
assert (image.get_objects(with_diff=False,
with_trun=False).shape[0] == 0)
def test_load_json_data(self):
image = Image.load_from_json_data(self.classes, self.data)
assert (image.width == 640 and image.height == 480)
assert (image.classes == ['A', 'B', 'C'])
assert (image.name == 'test_image')
objects_table = Table(
np.array([[0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 0, 0],
[2, 2, 2, 2, 2, 0, 0]]), self.columns)
assert (image.objects_table == objects_table)
def test_load_json_data(self):
image = Image.load_from_json_data(self.classes,self.data)
assert(image.width == 640 and image.height == 480)
assert(image.classes == ['A','B','C'])
assert(image.name == 'test_image')
objects_table = Table(np.array([
[0,0,0,0,0,0,0],
[1,1,1,1,1,0,0],
[2,2,2,2,2,0,0]]), self.columns)
assert(image.objects_table == objects_table)
def test_get_det_gt(self):
image = Image.load_from_json_data(self.classes,self.data)
objects_table = Table(np.array([
[0,0,0,0,0,0,0],
[1,1,1,1,1,0,0],
[2,2,2,2,2,0,0]]), self.columns)
assert(image.get_objects() == objects_table)
data = self.data.copy()
data['objects'][0]['diff'] = 1
data['objects'][1]['trun'] = 1
image = Image.load_from_json_data(self.classes,data)
objects_table = Table(np.array([
[0,0,0,0,0,1,0],
[1,1,1,1,1,0,1],
[2,2,2,2,2,0,0]]), self.columns)
assert(image.get_objects(with_diff=True,with_trun=True) == objects_table)
objects_table = Table(np.array([
[1,1,1,1,1,0,1],
[2,2,2,2,2,0,0]]),self.columns)
assert(image.get_objects(with_diff=False,with_trun=True) == objects_table)
# this should be default behavior
assert(image.get_objects() == objects_table)
objects_table = Table(np.array([
[2,2,2,2,2,0,0]]),self.columns)
assert(image.get_objects(with_diff=False,with_trun=False) == objects_table)
objects_table = Table(np.array([
[0,0,0,0,0,1,0],
[2,2,2,2,2,0,0]]), self.columns)
assert(image.get_objects(with_diff=True,with_trun=False) == objects_table)
# What if everything is filtered out?
data['objects'] = data['objects'][:-1]
objects_table = Table(np.array([
[0,0,0,0,0,1,0],
[1,1,1,1,1,0,1]]), self.columns)
image = Image.load_from_json_data(self.classes,data)
assert(image.get_objects(with_diff=True,with_trun=True) == objects_table)
assert(image.get_objects(with_diff=False,with_trun=False).shape[0] == 0)
def test_get_cls_counts_and_gt(self):
data = self.data.copy()
image = Image.load_from_json_data(self.classes, data)
assert (np.all(image.get_cls_counts() == np.array([1, 1, 1])))
assert (np.all(image.get_cls_gt() == np.array([True, True, True])))
assert (image.contains_class('A') == True)
assert (image.contains_class('B') == True)
data['objects'][0]['class'] = 'B'
image = Image.load_from_json_data(self.classes, data)
# doesn't actually have to be Series, can be ndarray for comparison
assert (np.all(image.get_cls_counts() == np.array([0, 2, 1])))
assert (np.all(image.get_cls_gt() == np.array([False, True, True])))
assert (image.contains_class('A') == False)
assert (image.contains_class('B') == True)
data['objects'] = []
image = Image.load_from_json_data(self.classes, data)
assert (np.all(image.get_cls_counts() == np.array([0, 0, 0])))
assert (np.all(image.get_cls_gt() == np.array([False, False, False])))
assert (image.contains_class('A') == False)
assert (image.contains_class('B') == False)
def test_get_cls_counts_and_gt(self):
data = self.data.copy()
image = Image.load_from_json_data(self.classes,data)
assert(np.all(image.get_cls_counts() == np.array([1,1,1])))
assert(np.all(image.get_cls_gt() == np.array([True,True,True])))
assert(image.contains_class('A') == True)
assert(image.contains_class('B') == True)
data['objects'][0]['class'] = 'B'
image = Image.load_from_json_data(self.classes,data)
# doesn't actually have to be Series, can be ndarray for comparison
assert(np.all(image.get_cls_counts() == np.array([0,2,1])))
assert(np.all(image.get_cls_gt() == np.array([False,True,True])))
assert(image.contains_class('A') == False)
assert(image.contains_class('B') == True)
data['objects'] = []
image = Image.load_from_json_data(self.classes,data)
assert(np.all(image.get_cls_counts() == np.array([0,0,0])))
assert(np.all(image.get_cls_gt() == np.array([False,False,False])))
assert(image.contains_class('A') == False)
assert(image.contains_class('B') == False)
def load_from_pascal(self, name, force=False):
"""
Look up the filename associated with the given name.
Read image names from provided filename, and construct a dataset from the
corresponding .xml files.
Save self to disk when loaded for caching purposes.
If force is True, does not look for cached data when loading.
"""
tt = ut.TicToc().tic()
print("Dataset: %s" % name),
filename = config.get_cached_dataset_filename(name)
if opexists(filename) and not force:
with open(filename) as f:
cached = cPickle.load(f)
self.classes = cached.classes
self.images = cached.images
print("...loaded from cache in %.2f s" % tt.qtoc())
return
print("...loading from scratch")
filename = config.pascal_paths[name]
self.classes = config.pascal_classes
with open(filename) as f:
imgset = [line.strip() for line in f.readlines()]
for i, img in enumerate(imgset):
tt.tic('2')
if tt.qtoc('2') > 2:
print(" on image %d/%d" % (i, len(imgset)))
tt.tic('2')
if len(img) > 0:
xml_filename = opjoin(config.VOC_dir, 'Annotations',
img + '.xml')
self.images.append(
Image.load_from_pascal_xml_filename(
self.classes, xml_filename))
filename = config.get_cached_dataset_filename(name)
print(" ...saving to cache file")
with open(filename, 'w') as f:
cPickle.dump(self, f)
print(" ...done in %.2f s\n" % tt.qtoc())
def load_from_pascal(self, name, force=False):
"""
Look up the filename associated with the given name.
Read image names from provided filename, and construct a dataset from the
corresponding .xml files.
Save self to disk when loaded for caching purposes.
If force is True, does not look for cached data when loading.
"""
tt = ut.TicToc().tic()
print("Dataset: %s" % name),
filename = config.get_cached_dataset_filename(name)
if opexists(filename) and not force:
with open(filename) as f:
cached = cPickle.load(f)
self.classes = cached.classes
self.images = cached.images
print("...loaded from cache in %.2f s" % tt.qtoc())
return
print("...loading from scratch")
filename = config.pascal_paths[name]
self.classes = config.pascal_classes
with open(filename) as f:
imgset = [line.strip() for line in f.readlines()]
for i, img in enumerate(imgset):
tt.tic("2")
if tt.qtoc("2") > 2:
print(" on image %d/%d" % (i, len(imgset)))
tt.tic("2")
if len(img) > 0:
xml_filename = opjoin(config.VOC_dir, "Annotations", img + ".xml")
self.images.append(Image.load_from_pascal_xml_filename(self.classes, xml_filename))
filename = config.get_cached_dataset_filename(name)
print(" ...saving to cache file")
with open(filename, "w") as f:
cPickle.dump(self, f)
print(" ...done in %.2f s\n" % tt.qtoc())
def test_get_whole_image_bbox(self):
image = Image(20, 10, [], 'test_image')
assert (image.get_whole_image_bbox() == BoundingBox((0, 0, 20, 10)))
image = Image(2, 100, [], 'test_image')
assert (image.get_whole_image_bbox() == BoundingBox((0, 0, 2, 100)))
def test_constructor(self):
image = Image(20, 10, ['A', 'B', 'C'], 'test_image')
assert (image.width == 20 and image.height == 10)
assert (image.classes == ['A', 'B', 'C'])
assert (image.name == 'test_image')
def test_get_whole_image_bbox(self): image = Image(20,10,[],'test_image') assert(image.get_whole_image_bbox() == BoundingBox((0,0,20,10))) image = Image(2,100,[],'test_image') assert(image.get_whole_image_bbox() == BoundingBox((0,0,2,100)))
slices = []
for i in range(num_bins):
hist = np.matrix(count_histogram_for_slice(assignments, im_width, im_height, \
num_bins, i, num_words)[1])
hist_sum = float(np.sum(hist))
slices.append(hist / hist_sum)
return slices
if __name__ == '__main__':
assignments = np.zeros((25, 3))
ind = 0
for i in range(5):
for j in range(5):
ass = 1
if (i, j) == (1, 1) or (i, j) == (2, 1) or (i, j) == (3, 1):
ass = 2
if (i, j) == (1, 2) or (i, j) == (2, 2) or (i, j) == (3, 2):
ass = 3
if (i, j) == (1, 3) or (i, j) == (2, 3) or (i, j) == (3, 3):
ass = 4
assignments[ind, :] = np.matrix([[i, j, ass]])
ind += 1
image = Image(size=(5, 5))
num_words = 4
slices = extract_horiz_sclices(3, assignments, image, num_words)
corr_stack = np.matrix([[7, 3, 0, 0], [2, 0, 3, 0], [7, 0, 0, 3]])
slice_stack = np.vstack(slices)
assert (corr_stack.all() == slice_stack.all())
#print slices
