class ActiveLearner:
def __init__(self, email, password, api_endpoint, project_id):
kauth = KiliAuth(email, password, api_endpoint=api_endpoint)
self.playground = Playground(kauth)
self.project_id = project_id
def get_assets_to_evaluate(self):
assets = self.playground.get_assets(project_id=self.project_id)
assets_to_evaluate = []
for asset in assets:
if len(asset['labels']) == 0:
assets_to_evaluate.append(asset)
return assets_to_evaluate
def prioritize_assets(self, assets, scorer, *args, **kwargs):
assets_score = [scorer(asset, *args, **kwargs) for asset in assets]
ranked_assets_with_score = sorted(list(zip(assets, assets_score)),
key=lambda x: x[1],
reverse=True)
ranked_assets = [
asset_with_score[0]
for asset_with_score in ranked_assets_with_score
]
return ranked_assets
def update_assets_priority(self, assets):
for i, asset in enumerate(tqdm(assets)):
asset_id = asset['id']
self.playground.update_properties_in_asset(asset_id=asset_id,
priority=i)
return True
def main(api_endpoint):
email = input('Enter email: ')
password = getpass.getpass()
source_project_id = input(
'Enter project IDs (separate them by "," if you want to provide several): '
)
kauth = KiliAuth(email=email, password=password, api_endpoint=api_endpoint)
playground = Playground(kauth)
df = pd.DataFrame(columns=['Project', 'Date', 'Email'])
for project_id in source_project_id.split(','):
project = playground.get_project(project_id=project_id)
assets = playground.get_assets(project_id=project_id)
title = project['title']
for asset in assets:
for label in asset['labels']:
created_at = label['createdAt'][:10]
author_email = label['author']['email']
df = df.append(
{
'Project': title,
'Date': created_at,
'Email': author_email
},
ignore_index=True)
df_grouped = df.groupby(['Project', 'Date', 'Email']).size()
time = datetime.now().strftime('%Y%m%d%H%M')
df_grouped.to_excel(f'labeler-stats-{time}.xlsx')
class TransferLearning:
def __init__(self, email, password, api_endpoint, project_id, minimum_number_of_assets_to_launch_training=100):
kauth = KiliAuth(email, password, api_endpoint=api_endpoint)
self.playground = Playground(kauth)
self.project_id = project_id
self.current_training_number = 0
self.last_training_number = -1
self.assets_seen_in_training = []
self.minimum_number_of_assets_to_launch_training = minimum_number_of_assets_to_launch_training
def _current_training_number(self):
return self.current_training_number
def get_assets_to_train(self):
assets = self.playground.get_assets(project_id=self.project_id)
assets_to_train = []
for asset in assets:
default_labels = get_labels_of_types(asset, ['DEFAULT'])
review_labels = get_labels_of_types(asset, ['REVIEWED'])
if len(review_labels) > 0:
asset['labels'] = [review_labels[-1]]
assets_to_train.append(asset)
elif len(default_labels) == 1:
asset['labels'] = [default_labels[-1]]
assets_to_train.append(asset)
elif len(review_labels) == 0 and len(default_labels) > 0:
print(
f'Asset {asset["id"]} has several labels: it should be reviewed')
else:
continue
return assets_to_train
def train(self, assets_to_train):
print(
f'Launch training for {len(assets_to_train)} assets: {[asset["id"] for asset in assets_to_train]}')
return
def launch_train(self):
time.sleep(SECONDS_TO_WAIT)
assets_to_train = self.get_assets_to_train()
if len(self.assets_seen_in_training) == 0:
filtered_assets_to_train = assets_to_train
else:
filtered_assets_to_train = [asset for asset in assets_to_train
if all([asset['id'] not in training
for training in self.assets_seen_in_training])]
if len(filtered_assets_to_train) >= self.minimum_number_of_assets_to_launch_training:
self.train(filtered_assets_to_train)
self.current_training_number += 1
self.assets_seen_in_training.append(
[asset['id'] for asset in filtered_assets_to_train])
def get_assets_to_predict(self):
assets = self.playground.get_assets(project_id=self.project_id)
assets_to_predict = []
for asset in assets:
labels = get_labels_of_types(asset, ['DEFAULT', 'REVIEWED'])
if len(labels) == 0:
assets_to_predict.append(asset)
return assets_to_predict
def predict(self, assets_to_predict):
print(
f'Launch inference for {len(assets_to_predict)} assets: {[asset["id"] for asset in assets_to_predict]}')
return
def launch_predict(self):
time.sleep(SECONDS_TO_WAIT)
if self.current_training_number == self.last_training_number:
print('Inference will not be launched for now...')
return
assets_to_predict = self.get_assets_to_predict()
if len(assets_to_predict) > 0:
current_training_number = self.current_training_number
self.predict(assets_to_predict)
self.last_training_number = current_training_number
def launch_tensorboard(self):
print('Starting Tensorboard...')
subprocess.Popen(['tensorboard', '--logdir=runs'])
print('You can access Tensorboard at http://localhost:6006\n')
def launch(self):
self.launch_tensorboard()
while True:
self.launch_train()
self.launch_predict()
import getpass
import json
import yaml
from tqdm import tqdm
from kili.authentication import KiliAuth
from kili.playground import Playground
email = input('Enter email: ')
password = getpass.getpass()
project_id = input('Enter project id: ')
kauth = KiliAuth(email=email, password=password)
playground = Playground(kauth)
assets = playground.get_assets(project_id=project_id)
for asset in tqdm(assets):
playground.delete_from_dataset(asset_id=asset['id'])
