def main(args):
data_dir_1 = args.data_dir_1
data_dir_2 = args.data_dir_2
keyword = args.keyword
name_data_1 = args.name_1
name_data_2 = args.name_2
output_dir = args.output_dir
container = EmbeddingContainer(name=name_data_1)
container.load(data_dir_1)
db_container = EmbeddingContainer(name=name_data_2)
db_container.load(data_dir_2)
container.merge(db_container, merge_key=keyword)
db_container.clear()
print(container)
print(container.DataFrame)
container.save(output_dir)
def create_embedding_container_from_featobj(folder_path, verbose=True):
"""Directly load feature object into embedding container.
Args:
folder_path: string, path to the folder of FeatureObject
verbose: Boolean, show the size of feature object if set True
Return:
container: EmbeddingContainer
"""
feature_importer = FeatureObject()
feature_importer.load(folder_path)
embeddings = feature_importer.embeddings
filenames = feature_importer.filename_strings
instance_ids = feature_importer.instance_ids
labels = feature_importer.label_ids
label_names = feature_importer.label_names
probabilities = feature_importer.probabilities
has_label_name = True if label_names is not None else False
# TODO
has_prob = True if probabilities is not None else False
# pseudo instance_ids
pseudo_instance_ids = np.arange(embeddings.shape[0])
if instance_ids is None or instance_ids.size == 0:
instance_ids = pseudo_instance_ids
num_feature, dim_feature = embeddings.shape
if verbose:
print('{} features with dim-{} are loaded'.format(
num_feature, dim_feature))
# err handling: label_ids.shape == 0
container = EmbeddingContainer(embedding_size=dim_feature,
prob_size=0,
container_size=num_feature)
if not has_label_name:
for inst_id, feat, label in zip(instance_ids, embeddings, labels):
# use filename_string as instance_id, convert to integer
#for postfix in ['.png', '.jpg', '.jpeg', '.JPG']:
# fn = fn.replace(postfix, '')
#pseudo_instance_id = int(fn)
inst_id = int(inst_id)
container.add(inst_id, label, feat)
else:
for inst_id, feat, label, name in zip(instance_ids, embeddings, labels,
label_names):
inst_id = int(inst_id)
container.add(inst_id, label, feat, label_name=name)
return container
def main(args):
data_dir = args.data_dir
out_dir = args.out_dir
if data_dir is None:
raise ValueError('data_dir or database should be assigned.')
container = EmbeddingContainer()
container.load(data_dir)
hg = HierarchicalGrouping(container)
df = hg.auto_label_subgroup(label_id=1)
print(df)
def main(args):
data_dir = args.data_dir
out_dir = args.output_dir
container = EmbeddingContainer()
container.load(data_dir)
tsne = TSNE(container,
n_iter=args.iterations,
n_jobs=args.n_jobs,
perplexity=args.perplexity)
tsne.run()
tsne.save_fig(out_dir)
def main(args):
data_dir = args.data_dir
out_dir = args.out_dir
query_command = args.query_command
anchor_command = args.anchor_command
# TODO: sanity check
container = EmbeddingContainer()
result = ResultContainer()
container.load(data_dir)
command = '{}->{}'.format(query_command, anchor_command)
query_ids, anchor_ids = container.get_instance_id_by_cross_reference_command(command)
query_embeddings = container.get_embedding_by_instance_ids(query_ids)
anchor_embeddings = container.get_embedding_by_instance_ids(anchor_ids)
num_of_anchor = anchor_embeddings.shape[0]
num_of_query = query_embeddings.shape[0]
agent = IndexAgent(agent_type='HNSW',
instance_ids=anchor_ids,
embeddings=anchor_embeddings)
all_query_ids, all_retrieved_ids, all_retrieved_distances = [], [], []
with tqdm(total=num_of_query) as pbar:
for _idx, (query_id, qeury_emb) in enumerate(zip(query_ids, query_embeddings)):
retrieved_ids, retrieved_distances = agent.search(qeury_emb, top_k=num_of_anchor)
retrieved_ids = np.squeeze(retrieved_ids)
retrieved_distances = np.squeeze(retrieved_distances)
all_query_ids.extend(np.array(query_id).repeat(num_of_anchor))
all_retrieved_ids.extend(retrieved_ids)
all_retrieved_distances.extend(retrieved_distances)
pbar.update()
print('Indexing finished, {} retrieved events'.format(len(all_retrieved_ids)))
print('Start exporting results...')
start_time = time()
result._event_buffer = pd.DataFrame(
{
Fields.query_instance_id: all_query_ids,
Fields.retrieved_instance_id: all_retrieved_ids,
Fields.retrieved_distance: all_retrieved_distances,
}
)
result.save(out_dir)
print('Done, saving results take {} seconds.'.format(time() - start_time))
def __init__(self, embedding_size, prob_size, config_dict, mode='online'):
"""Evaluator Builder.
The object builds evaluation functions according to the given configuration
and manage shared data (embeddings, labels and attributes) in container objects.
Args:
embedding_size: Integer describes 1d embedding size.
prob_size: Integer describes size of the logits.
config_dict: Dict, loaded yaml foramt dict.
mode: String, `online` or `offline`.
Building procedure: (TODO @kv: update these steps)
* parse the config
* allocate containers
* create evaluations
* add datum
* run evaluate
* (optional) get update_ops
TODO:
- deprecate attribute container
"""
self.configs = ConfigParser(config_dict)
# allocate shared embedding containers
container_size = self.configs.container_size
self.embedding_size = embedding_size
self.prob_size = prob_size
self.embedding_container = EmbeddingContainer(embedding_size, prob_size, container_size)
self.mode = mode
if self.mode not in ['online', 'offline']:
raise ValueError('Evaluator mode: {} is not defined.'.format(self.mode))
self._build()
self._instance_counter = 0
self._total_metrics = {}
self._results = {}
# Allocate general query interface
if not self.configs.database[config_fields.database_type]:
# TODO @kv: consistent check with query condition
print('No attribute database')
self.query_interface = None
else:
self.query_interface = QueryInterface(self.configs.database)
print('Attribute database is initialized.')
def main(args):
if args.data_dir is None:
return
container = EmbeddingContainer()
container.load(args.data_dir)
all_embeddings = container.embeddings
instance_ids = container.instance_ids
all_embeddings = container.get_embedding_by_instance_ids(instance_ids)
agent = IndexAgent('HNSW', instance_ids, all_embeddings, distance_measure='ip')
print(container)
for label_id in container.label_ids:
same_class_inst_ids = container.get_instance_ids_by_label_ids(label_id)
same_class_embeddings = container.get_embedding_by_instance_ids(same_class_inst_ids)
num_inst_same_class = len(same_class_inst_ids)
retrieved_indexes, similarities = agent.search(
same_class_embeddings, top_k=2*num_inst_same_class, is_similarity=True)
break
def main(args):
data_dir = args.data_dir
out_dir = args.out_dir
if data_dir is None:
raise ValueError('data_dir or database should be assigned.')
feature_object = FeatureObject()
feature_object.load(data_dir)
embeddings = feature_object.embeddings
embeddings = np.squeeze(embeddings)
filename_strings = feature_object.filename_strings
label_ids = feature_object.label_ids
label_names = feature_object.label_names
instance_ids = np.arange(embeddings.shape[0])
# Push all embeddings into container
embedding_container = EmbeddingContainer(
embedding_size=embeddings.shape[1],
prob_size=0,
container_size=embeddings.shape[0])
for emb, inst_id, label_id in zip(embeddings, instance_ids, label_ids):
embedding_container.add(inst_id, label_id, emb)
manifold = Manifold(embedding_container, label_names)
centers = manifold.class_center()
c2c_matrix = manifold.center_to_center_relation()
c2all_relation = manifold.center_to_all_instance_relation()
for center_label, center_feature in centers.items():
manifold.distance_trace(center_label, center_feature, 200)
manifold.locality_analysis()
def main():
args = parser.parse_args()
config_path = args.config
data_type = args.data_type
data_dir = args.data_dir
out_dir = args.out_dir
anchor_database_dir = args.anchor_database
status = status_fields.not_determined
# check input is given
if not data_dir:
raise ValueError('data_dir must be assigned!')
if out_dir is not None:
if not os.path.exists(out_dir):
os.makedirs(out_dir)
# argument logic
if data_dir and anchor_database_dir:
status = status_fields.evaluate_query_anchor
elif data_dir and anchor_database_dir is None:
status = status_fields.evaluate_single_container
if not config_path:
# TODO: @kv: Generate the default config.
raise ValueError('evaluation configuration must be assigned!')
try:
with open(config_path, 'r') as fp:
config_dict = yaml.load(fp)
except:
raise IOError('Can not load yaml from {}.'.format(config_path))
# TODO: create default config instead of error.
# Prepare data container
container = None
for case in switch(status):
print('{} Executes {}'.format(APP_SIGNATURE, status))
if case(status_fields.evaluate_single_container):
container = EmbeddingContainer(name='single_container')
if data_type in ['embedding_container', 'embedding_db']:
container.load(data_dir)
# end of switch case
break
if case(status_fields.evaluate_query_anchor):
"""TODO: Use native method: merge()
1. Merge two containers
2. Add `query->anchor` command in cross_reference
3. Change number of database
"""
container = EmbeddingContainer(name='query')
anchor_container = EmbeddingContainer(name='anchor')
# load query
if data_type in ['embedding_container', 'embedding_db']:
container.load(data_dir)
# load anchor
if data_type in ['embedding_container', 'embedding_db']:
anchor_container.load(anchor_database_dir)
container.merge(anchor_container,
merge_key='merge_record',
label_id_rearrange=True)
# clear buffer
anchor_container.clear()
# Change config TODO: A little bit hacky, modify in future
# TODO: It seems not work well
_opt = config_fields.evaluation_options
_rank = 'RankingEvaluation'
_attr = config_fields.attribute
_cref = config_fields.cross_reference
_smp = config_fields.sampling
_cmd = 'merge_record.query -> merge_record.anchor'
config_dict[_opt][_rank][_attr][_cref] = list(
filter(None, config_dict[_opt][_rank][_attr][_cref]))
if _cmd not in config_dict[_opt][_rank][_attr][_cref]:
config_dict[_opt][_rank][_attr][_cref].append(_cmd)
config_dict[_opt][_rank][_smp][
'num_of_db_instance_per_class'] = 1000
# end of switch case
break
# Build and run evaluation
evaluator = EvaluatorBuilder(args.embedding_size,
args.prob_size,
config_dict,
mode='offline')
print(container)
evaluator.add_container(container)
evaluator.evaluate()
# Show Results
for eval_name, result_container in evaluator.results.items():
print(eval_name)
display_name = display_namemap[
eval_name] if eval_name in display_namemap else eval_name
reporter = ReportWriter(result_container)
overall_report = reporter.overall_report
print(overall_report)
if out_dir:
path = '/'.join([out_dir, 'result_{}'.format(display_name)])
result_container.save(path)
if status == status_fields.evaluate_query_anchor and out_dir:
path = '/'.join([out_dir, 'merged_container'])
container.save(path)