To kill celery task workers, use:

ps -ef | grep [YOUR_WORKER_NAME] | awk '{ print $2 }' | xargs kill -9

Do NOT grep celery or you may kill other tasks' workers.

By default, celery uses prefork worker pool, which is incompatible with pytorch non-forkable lib. So we switch to thread-based gevent pool. To install gevent:

conda install gevent

Put detection_state_dict.pth and score_state_dict.pth to data/.
If model was trained with DataParallel, you may need to re-save state_dict with extract_state_dict.ipynb.

Prepare celery-config.py and config.py.
celery-config.py:

broker_url = 'redis://:[REDIS_PASSWORD]@r-wz9j4vwobcb81ws9mapd.redis.rds.aliyuncs.com:6379/0'
result_backend = 'redis://:[REDIS_PASSWORD]@r-wz9j4vwobcb81ws9mapd.redis.rds.aliyuncs.com:6379/0'
device_name = 'cpu' # or 'cuda:0' if you want to use gpu

config.py:

access_key_id = 'xxxxx'
access_key_secret = 'yyyyy'

start celery worker:

celery -A root-worker worker --pool gevent --concurrency 1 --time-limit 60

or with nohup:

nohup celery -A root-worker worker --pool gevent --concurrency 1 --time-limit 60 > worker.log &