Pytorch model prediction in production with uwsgi - nginx

Question

I have a problem deploying a pytorch model in production. For a demonstration, I build a simple model and a flask app. I put everything in a docker container (pytorch+flask+uwsgi) plus another container for nginx. Everything is running well, my app is rendered and I can navigate inside. However, well I navigate into the URL that launches a prediction of the model, the server hangs and does not seem to compute anything.
The uWSGI is run like this:
/opt/conda/bin/uwsgi --ini /usr/src/web/uwsgi.ini
with uwsgi.ini
[uwsgi]
#application's base folder
chdir = /usr/src/web/
#python module to import
wsgi-file = /usr/src/web/wsgi.py
callable = app
#socket file's location
socket = /usr/src/web/uwsgi.sock
#permissions for the socket file
chmod-socket = 666
# Port to expose
http = :5000
# Cleanup the socket when process stops
vacuum = true
#Log directory
logto = /usr/src/web/app.log
# minimum number of workers to keep at all times
cheaper = 2
processes = 16
As said, the server hangs and I finally got a timeout. What is strange is when I run the flask application directly (also in the container) with
python /usr/src/web/manage.py runserver --host 0.0.0.0
I get my prediction in no time

Answer 1

I think this is related to
https://discuss.pytorch.org/t/basic-operations-do-not-work-in-1-1-0-with-uwsgi-flask/50257
Maybe try as mentioned there:
app = flask.Flask(__name__)
segmentator = None
#app.before_first_request
def load_segmentator():
global segmentator
segmentator = Segmentator()
where Segmentator is a class with pytorch’s nn.Module, which loads weights in __init__
FYI this solution worked for me with one app but not the other