def create_iterative_training_classifier(current_user, base_classifier_id,
classifier_name, epoch, video_list):
base_classifier = Classifier.query.filter(
Classifier.id == base_classifier_id).first()
if not base_classifier:
return None
label_list = str(base_classifier.labels).split(',')
print 'create iterative training, labels are %s ' % str(label_list)
image_list_file_path, label_list_file_path, label_name_file_path = turkic_replacement.dump_image_and_label_files(
video_list, label_list, remove_none_frame=True)
classifier = Classifier.create(name=classifier_name,
owner_id=current_user.id)
# add videos
classifier.training_image_list_file_path = image_list_file_path
classifier.training_label_list_file_path = label_list_file_path
classifier.training_label_name_file_path = label_name_file_path
classifier.parent_id = base_classifier_id
classifier.model_name = classifier_name
classifier.epoch = epoch
classifier.network_type = config.NETWORK_TYPE_FASTER_RCNN
classifier.task_type = config.TASK_TYPE_TRAIN
# add these labels and videos to the classifier
classifier.videos = base_classifier.videos
for video_id in video_list:
video = Video.query.filter(Video.id == video_id).first()
if video:
classifier.videos.append(video)
classifier.labels = ','.join(label_list)
# get id of the classifier
classifier_id = classifier.id
print 'generate iterative classifier with id %s ' % str(classifier_id)
# prepare training
classifier.training_start_time = int(time.time() * 1000)
train_set_name = os.path.splitext(
ntpath.basename(str(image_list_file_path)))[0]
base_image_name = util.get_classifier_image_name(base_classifier.name,
base_classifier.id)
result_image_name = util.get_classifier_image_name(classifier_name,
classifier_id)
task_id = launch_training_docker_task(base_image_name,
result_image_name,
classifier_id,
train_set_name,
epoch,
weights='iterative')
print 'launched the pretrain docker with task id %s ' % str(task_id)
classifier.task_id = task_id
db.session.commit()
def push_classifier(classifier, push_tag_name):
image_name = util.get_classifier_image_name(classifier.name, classifier.id)
client = docker.from_env()
try:
image = client.images.get(str(image_name))
except docker.errors.ImageNotFound as e:
logger.error(e)
return 'Error: no contaienr image found for Classifier: {}'.format(
classifier.name)
push_image_task.delay(image_name, push_tag_name)
return 'Success: The image is being pushed to registery. \n You can pull the image with command: docker pull {0}:{1} once is finished. You can go to {0} to check its status'.format(
config.CONTAINER_REGISTRY_URL, str(push_tag_name))
def create_short_running_test_classifier(base_classifier_id, time_remains=100):
base_classifier = Classifier.query.filter(
Classifier.id == base_classifier_id).first()
if not base_classifier:
return None
host_port = util.get_available_port()
base_image_name = util.get_classifier_image_name(base_classifier.name,
base_classifier.id)
fake_classifier_id = 'fake-' + str(base_classifier_id)
task_id = launch_test_docker_task(fake_classifier_id, base_image_name,
time_remains, host_port)
print 'launched the test docker with task id %s ' % str(task_id)
return host_port
def create_test_classifier(current_user,
base_classifier_id,
time_remains=1000):
base_classifier = Classifier.query.filter(
Classifier.id == base_classifier_id).first()
if not base_classifier:
return None
classifier = Classifier.create(name=base_classifier.name,
owner_id=current_user.id)
# add videos
classifier.training_image_list_file_path = base_classifier.training_image_list_file_path
classifier.training_label_list_file_path = base_classifier.training_label_list_file_path
classifier.training_label_name_file_path = base_classifier.training_label_name_file_path
classifier.task_type = config.TASK_TYPE_TEST
classifier.parent_id = base_classifier_id
classifier.model_name = base_classifier.model_name
classifier.epoch = base_classifier.epoch
classifier.network_type = base_classifier.network_type
# add these labels and videos to the classifier
classifier.videos = base_classifier.videos
classifier.labels = base_classifier.labels
# get id of the classifier
classifier_id = classifier.id
print 'generate classifier with id %s ' % str(classifier_id)
classifier.training_start_time = int(time.time() * 1000)
host_port = util.get_available_port()
base_image_name = util.get_classifier_image_name(base_classifier.name,
base_classifier.id)
task_id = launch_test_docker_task(classifier_id, base_image_name,
time_remains, host_port)
print 'launched the test docker with task id %s ' % str(task_id)
classifier.task_id = task_id
classifier.image_id = task_id
classifier.container_id = task_id
db.session.commit()
return host_port
def delete_classifier(classifier_id):
'''
1. db entity
2. label_list, image_list, label_name
3. container, image
'''
classifier = Classifier.query.filter(
Classifier.id == classifier_id).first()
util.remove_file_if_exist(classifier.training_image_list_file_path)
util.remove_file_if_exist(classifier.training_label_list_file_path)
util.remove_file_if_exist(classifier.training_label_name_file_path)
container_image_name = util.get_classifier_image_name(
classifier.name, classifier.id)
client = docker.from_env()
try:
client.images.remove(container_image_name, force=True)
except docker.errors.ImageNotFound as e:
logger.info(
'contianer image {} not found'.format(container_image_name))
logger.info(e)
classifier.delete()
def run_onetime_classifier_test(base_classifier_id, input_image_path, min_cf):
base_classifier = Classifier.query.filter(
Classifier.id == base_classifier_id).first()
if not base_classifier:
return None
base_image_name = util.get_classifier_image_name(base_classifier.name,
base_classifier.id)
if not util.has_container_image(base_image_name):
return None
docker_name = 'onetime_test_classifier_' + base_classifier_id + '_' + str(
random.getrandbits(16))
output_image_path = '{}_output.jpg'.format(input_image_path)
ret_code = util.issue_blocking_cmd([
'nvidia-docker', 'run', '--rm', '-v', '/tmp:/tmp', '--name',
docker_name, base_image_name, 'tools/tpod_detect_cli.py',
'--input_image', input_image_path, '--min_cf', min_cf,
'--output_image', output_image_path
])
if ret_code:
# something went wrong
return None
return output_image_path
def create_evaluation(classifier_id, name, video_list):
classifier = Classifier.query.filter(
Classifier.id == classifier_id).first()
if not classifier:
print 'classifier not exist'
return None
label_list = classifier.labels.split(',')
evaluation_set = EvaluationSet(name=name)
for video_id in video_list:
video = Video.query.filter(Video.id == video_id).first()
evaluation_set.videos.append(video)
classifier.evaluation_sets.append(evaluation_set)
db.session.add(evaluation_set)
db.session.commit()
print 'created evaluation set with name %s id %s ' % (
str(name), str(evaluation_set.id))
evaluation_result_name = str(evaluation_set.id)
base_image_name = util.get_classifier_image_name(classifier.name,
classifier.id)
# prepare label data
image_list_file_path, label_list_file_path, label_name_file_path = turkic_replacement.dump_image_and_label_files(
video_list, label_list, remove_none_frame=True)
# note: this evaluation set name is the unique name to indicate the file name for the label, video,
# not the name of the data entity
dataset_path = util.get_dataset_path()
eval_path = util.get_eval_path()
evaluation_set_name = os.path.splitext(
ntpath.basename(str(image_list_file_path)))[0]
evaluation_task.apply_async(
(dataset_path, eval_path, classifier_id, base_image_name,
evaluation_set_name, evaluation_result_name))