def load_vocab_from_path(vocabs_file, dataset_path):
# there should only be a single hit
vocab_file_list = glob.glob(str(vocabs_file))
try:
assert len(vocab_file_list) == 1
except AssertionError:
logging.error(
'Something is wrong. The dataset paths are probably wrong, or the basename in '
'load_vocab_from_path does not match any vocab .csv file.')
logging.error('basename: ' + str(vocabs_file))
logging.error('vocab_file_list: ' + str(vocab_file_list))
exit()
vocab_file = open(vocab_file_list[0], 'r')
reader = csv.reader(vocab_file, delimiter=',')
des_file_list = glob.glob(dataset_path + '/*descriptors.csv')
# i will iterate on the files in des_file_list
i = 0
y = []
X = []
for xi in reader:
xi = list(map(lambda x: float(x), xi))
X.append(xi)
y.append(float(infer_label(des_file_list[i])))
#print(des_file_list[i])
i += 1
#print(xi)
#print(infer_label(des_file_list[i]))
#print(y)
return X, y
def generate_vocabulary(dataset_path, size, model=None):
# set up output variables
y = []
X = list()
num_des_list = list(
) # list that mantains the number of descriptors in each video
# set up clustering model
if (model == None):
model = KMeans(n_clusters=size, tol=1e-2)
descriptor_file_list = glob.glob(dataset_path + '*descriptors.csv')
logging.info('Loaded ' + str(len(descriptor_file_list)) +
' descriptors files for vocabulary generation.')
_ = []
for descriptor_file in descriptor_file_list:
prev_X_len = len(X)
X, _ = load_samples_from_file(descriptor_file, X, _)
yi = infer_label(descriptor_file)
y.append(yi)
num_des_list.append(len(X) - prev_X_len)
# cluster the data, and get labels
labels = model.fit_predict(X)
#labels = model.predict(X)
logging.info('Done clustering.')
num_labels = len(set(labels)) # figure out the number of unique labels
video_dicts = [[]]
video_dicts[0] = [0 for _ in range(0, num_labels)]
k = 0 # iterates on the number of videos
j = 0 # iterates on the number of descriptors
for i in range(0, len(labels)):
video_dicts[k][labels[i]] += 1
if j == num_des_list[k]:
j = 0
k += 1
video_dicts.append([0 for _ in range(0, num_labels)])
j += 1
if DUMP_TO_FILE == True:
f = open(
Constants.DATA_DIR +
'/vocabs/scovanner/python/kmeans-eucl-py-yash-' + str(size) +
'.csv', 'w')
writer = csv.writer(f)
for dict in video_dicts:
writer.writerow(dict)
return video_dicts, y
def generate_ngram_vocab(dataset_path,
feature_path,
size,
model=None,
n=2,
ngram_strategy='naive',
window_size=3):
# set up output variables
y = []
X = list()
num_des_list = list(
) # list that mantains the number of descriptors in each video
# set up clustering model
if (model == None):
model = KMeans(n_clusters=size, tol=1e-2)
descriptor_file_list = glob.glob(dataset_path + '*descriptors.csv')
feature_file_list = glob.glob(feature_path + '*features.csv')
logging.info('Loaded ' + str(len(descriptor_file_list)) +
' files for vocab generation.')
_ = []
i = 0
X_temp = []
F_temp = []
for descriptor_file in descriptor_file_list:
feature_file = feature_file_list[i]
prev_X_len = len(X)
X_temp, F_temp, _ = load_descriptors_and_features(
descriptor_file, feature_file, X_temp, F_temp, _)
yi = infer_label(descriptor_file)
y.append(yi)
for j in range(0, len(X_temp)):
F_temp[j].extend(X_temp[j])
F_temp.sort(key=lambda f: int(f[2]))
#for fi in F_temp:
# del fi[:3]
X.extend(F_temp)
num_des_list.append(len(X) - prev_X_len)
i += 1
X_temp = []
F_temp = []
num_des_list[-1] -= (n + 1)
# cluster the data, and get labels
labels = model.fit_predict([xi[:3] for xi in X])
logging.info('Done clustering.')
j = 0 # iterates on the number of descriptors in a video
k = 0 # iterates on the number of videos
ngram_index_string = ''
video_dicts = []
ngram_dict = init_ngram_dict(size, n)
empty_ngram_dict = ngram_dict
for des_i in range(0, len(labels) - n):
xi = X[des_i]
xi_t = int(xi[2])
ngram_index_string = str(labels[des_i]) + '-'
if ngram_strategy == 'naive':
for ngram_i in range(1, n):
ngram_index_string += str(labels[des_i + ngram_i]) + '-'
if ngram_index_string in ngram_dict:
ngram_dict[str(ngram_index_string)] += 1
else:
ngram_dict[str(ngram_index_string)] = 1
if ngram_strategy == 'window':
if n != 2:
logging.error('Unsupported strategy for n > 2')
return
for w in range(des_i + 1, len(labels) - n):
ngram_index_string = str(labels[des_i]) + '-'
xw = X[w]
xw_t = int(xw[2])
if abs(xi_t - xw_t) > window_size:
break
ngram_index_string += str(labels[w]) + '-'
if ngram_index_string in ngram_dict:
ngram_dict[str(ngram_index_string)] += 1
else:
ngram_dict[str(ngram_index_string)] = 1
if j == num_des_list[k]:
j = 0
k += 1
ngram_dict = sorted(ngram_dict.items(), key=operator.itemgetter(0))
#print(ngram_dict)
ngram_dict = [
num_word_occurrences
for (key, num_word_occurrences) in ngram_dict
]
if len(ngram_dict) < np.power(size, n):
ngram_dict.extend(
[0 for i in range(0,
np.power(size, n) - len(ngram_dict))])
video_dicts.append(ngram_dict)
ngram_dict = empty_ngram_dict.copy()
j += 1
if DUMP_TO_FILE == True:
f = open(
Constants.DATA_DIR + '/vocabs/bbrister-ngrams/bbrister-window-' +
str(size) + '-ngram' + str(n) + '.csv', 'w')
writer = csv.writer(f)
for dict in video_dicts:
writer.writerow(dict)
return video_dicts, y
def generate_vocabulary_2d_quadrants(dataset_path,
features_path,
size_words,
size_angles,
model=None,
vocabs_path=None,
quadrant='x'):
# input checking
if size_angles != 2 and size_angles != 8:
logging.error('Unsupported number of quadrants')
sys.exit()
if size_angles == 2 and quadrant != 'x' and quadrant != 'y' and quadrant != 'z':
logging.error('2 quadrants case: unsupported \'quadrant\' parameter')
sys.exit()
feature_file_list = glob.glob(features_path + '*features.csv')
descriptor_file_list = glob.glob(dataset_path + '*descriptors.csv')
logging.info('Loaded ' + str(len(descriptor_file_list)) +
' descriptors files for vocabulary generation.')
logging.info('Loaded ' + str(len(feature_file_list)) +
' features files for vocabulary generation.')
i = 0
F = []
X = []
y = []
num_des_list = []
# load descriptors, features and true video labels
for feature_file in feature_file_list:
prev_X_len = len(X)
descriptor_file = descriptor_file_list[i]
X, F, y = load_descriptors_and_features(descriptor_file, feature_file,
X, F, y)
yi = infer_label(descriptor_file)
y.append(yi)
num_des_list.append(len(X) - prev_X_len)
i += 1
assert len(X) == len(F)
# perform clustering on descriptors
logging.info('Clustering descriptors...')
if model == None:
model = KMeans(n_clusters=size_words)
word_labels = model.fit_predict(X)
# perform quadrant checking on angles
logging.info('Clustering angles...')
F = np.array(F)[:, 3:6]
angle_labels = []
for angle_list in F:
x = float(angle_list[0])
y = float(angle_list[1])
z = float(angle_list[2])
angle_labels.append(
get_angle_quadrant_label(x, y, z, size_angles, quadrant))
logging.info('Building 2d signatures...')
video_2d_dict = np.zeros((size_words, size_angles), dtype=np.int)
video_dicts = []
if size_angles == 2:
f = open(
Constants.DATA_DIR +
'/vocabs/bbrister-quadrants/bbrister-quadrants-' + str(quadrant) +
'-' + str(size_angles) + '-size-' + str(size_words) + '.csv', 'w')
if size_angles == 8:
f = open(
Constants.DATA_DIR +
'/vocabs/bbrister-quadrants/bbrister-quadrants-' +
str(size_angles) + '-size-' + str(size_words) + '.csv', 'w')
writer = csv.writer(f)
j = 0
k = 0
for i in range(0, len(word_labels)):
video_2d_dict[word_labels[i]][angle_labels[i]] += 1
if j == num_des_list[k] - 1:
j = 0
k += 1
video_dicts.append(video_2d_dict)
if DUMP_TO_FILE == True:
writer.writerow(video_2d_dict.flatten())
video_2d_dict = np.zeros((size_words, size_angles), dtype=np.int)
else:
j += 1
return video_dicts, y
def generate_vocabulary_2d(dataset_path,
features_path,
size_words,
size_angles,
model=None,
vocabs_path=None):
feature_file_list = glob.glob(features_path + '*features.csv')
descriptor_file_list = glob.glob(dataset_path + '*descriptors.csv')
logging.info('Loaded ' + str(len(descriptor_file_list)) +
' descriptors files for vocabulary generation.')
logging.info('Loaded ' + str(len(feature_file_list)) +
' features files for vocabulary generation.')
i = 0
F = []
X = []
y = []
num_des_list = []
# load descriptors, features and true video labels
for feature_file in feature_file_list:
prev_X_len = len(X)
descriptor_file = descriptor_file_list[i]
X, F, y = load_descriptors_and_features(descriptor_file, feature_file,
X, F, y)
yi = infer_label(descriptor_file)
y.append(yi)
num_des_list.append(len(X) - prev_X_len)
i += 1
assert len(X) == len(F)
# perform clustering on descriptors
logging.info('Clustering descriptors...')
if model == None:
model = KMeans(n_clusters=size_words)
word_labels = model.fit_predict(X)
# perform clustering on angles
logging.info('Clustering angles...')
F = np.array(F)
model = KMeans(n_clusters=size_angles)
# only take the angles' values
angle_labels = model.fit_predict(F[:, 3:6])
logging.info('Building 2d signatures...')
video_2d_dict = np.zeros((size_words, size_angles), dtype=np.int)
video_dicts = []
f = open(
Constants.DATA_DIR + '/vocabs/bbrister-angles/bbrister-angles-' +
str(size_angles) + '-size-' + str(size_words) + '.csv', 'w')
writer = csv.writer(f)
j = 0
k = 0
for i in range(0, len(word_labels)):
video_2d_dict[word_labels[i]][angle_labels[i]] += 1
if j == num_des_list[k] - 1:
j = 0
k += 1
video_dicts.append(video_2d_dict)
if DUMP_TO_FILE == True:
writer.writerow(video_2d_dict.flatten())
video_2d_dict = np.zeros((size_words, size_angles), dtype=np.int)
else:
j += 1
return video_dicts, y
def generate_vocabulary_with_pruning(dataset_path,
features_path,
size,
model=None,
scale=1.6,
octave=0,
same_scale=False):
# set up output variables
y = []
X = list()
num_des_list = list(
) # list that mantains the number of descriptors in each video
# set up clustering model
if (model == None):
model = KMeans(n_clusters=size, tol=1e-2)
descriptor_file_list = glob.glob(dataset_path + '*descriptors.csv')
feature_file_list = glob.glob(features_path + '*features.csv')
logging.info('Loaded ' + str(len(descriptor_file_list)) +
' descriptors files for vocabulary generation.')
logging.info('Loaded ' + str(len(feature_file_list)) +
' features files for vocabulary generation.')
i = 0
F = []
X = []
y = []
for descriptor_file in descriptor_file_list:
feature_file = feature_file_list[i]
prev_X_len = len(X)
X, F, y = load_samples_from_file_same_dim(descriptor_file,
feature_file, X, F, y,
octave, scale)
yi = infer_label(descriptor_file)
y.append(yi)
num_des_list.append(len(X) - prev_X_len)
i += 1
logging.info('Loaded ' + str(len(X)) +
' descriptors after octave/scale pruning')
# cluster the data, and get labels
model.fit(X)
labels = model.predict(X)
logging.info('Done clustering.')
num_labels = len(set(labels)) # figure out the number of unique labels
video_dicts = [[]]
video_dicts[0] = [0 for _ in range(0, num_labels)]
k = 0 # iterates on the number of videos
j = 0 # iterates on the number of descriptors
fi = open(
Constants.DATA_DIR +
'/vocabs/bbrister-samedim/bbrister-samedim-octave' + str(octave) +
'-size-' + str(size) + '.csv', 'w')
writer = csv.writer(fi)
for i in range(0, len(labels)):
video_dicts[k][labels[i]] += 1
if j == num_des_list[k] - 1:
j = 0
k += 1
video_dicts.append([0 for _ in range(0, num_labels)])
if DUMP_TO_FILE:
writer.writerow(video_dicts[k - 1])
else:
j += 1
return video_dicts, y