def getObjectMetric(pair, metric_name):
if metric_name == 'self':
return pair['object']
if metric_name == '__path__':
return getObjectPath(pair['object'])
annotations = pair['annotations']
for a in annotations:
if metric_name in a:
return a[metric_name]
def getAnnotations(object, all_annotations_files):
object_file_path = getObjectPath(object)
annotations_files_paths_prefix = getAnnotationsFilesPrefix(object_file_path)
possible_annotations_files_paths = [
annotations_files_paths_prefix + '.txt',
annotations_files_paths_prefix + '_left.txt',
annotations_files_paths_prefix + '_right.txt'
]
annotations_files = [{
'path': path,
'content': all_annotations_files[path]
} for path in filter(lambda p: p in all_annotations_files, possible_annotations_files_paths)]
extracted_annotations = map(lambda c: extractAnnotations(object, c), annotations_files)
annotations = filter(lambda a: a != None, extracted_annotations)
return list(annotations)
def countVideosInClasses(pairs, get_classes_function):
videos_by_classes = {}
for pair in tqdm(filter(isFrame, pairs),
desc='Counting videos in classes'):
classes = get_classes_function(annotation['emotion'])
for class_name in classes:
if not (class_name in videos_by_classes):
videos_by_classes[class_name] = {}
video_path = getObjectPath(pair['object'])
if not (video_path in videos_by_classes[class_name]):
videos_by_classes[class_name][video_path] = True
number_of_videos_by_classes = {}
for class_name in videos_by_classes:
number_of_videos_by_classes[class_name] = len(
videos_by_classes[class_name])
return number_of_videos_by_classes
def printPairsPaths(pairs):
for p in pairs:
print(getObjectPath(p['object']))
dataset_specific_api = getDatasetSpecificApi('experiment')
pairs = Pairs(dataset_dir,
get_classes_function=dataset_specific_api.getClasses)
# pairs.filterBy({'__class__': ['happy']})
pairs.setReturnNames(True)
hogs_and_pairs = list(getPairsMetricsWithPairs(pairs, 'hog', threads=24))
pairs.setReturnNames(False)
hogs = np.array([e[0] for e in hogs_and_pairs])
print(hogs.shape[0], 'elements')
kmeans = KMeans(n_clusters=clusters_number, random_state=0).fit(hogs)
labels = kmeans.labels_
ids = np.array([e[1] for e in hogs_and_pairs])
result = {
l: [ids[i] for i in np.where(labels == l)][0]
for l in range(clusters_number)
}
for cluster_id in result:
print(cluster_id)
for object_id in result[cluster_id]:
object = pairs[object_id]['object']
print(getObjectPath(object))
# import torch
# from kmeans_pytorch import kmeans
# # kmeans
# cProfile.run("cluster_ids_x, cluster_centers = kmeans(X=torch.from_numpy(hogs), num_clusters=clusters_number, distance='cosine', device=torch.device('cuda'))")