UWSGI workers + Tensorflow - how to set up model initialization for each worker?

Question

Python

I ran into a problem with Flask and Tensorflow in a REST API service.
Given:
Let's say we have a web service written with Flask that is served on UWSGI. The service also uses a keras model based on the Tensorflow backend.
The service has a global variable (an initialized model) that must be initialized when the worker starts.
Issues: The problem is that the initialization of the model must occur at the start of each service. For example, in the wsgi.py file, after importing the application, we initialize the model like this:

from web.app import application
from keras_backend import KerasModel

application.keras_model = KerasModel(model_path='models/keras-weights.hdf5')

In this approach, the initialization of the model occurs once at the start of the very first worker. And accordingly it is stored only there.
On the yuvzgi config with one thread and one process, everything works. If you make the value of the workers option (they are processes) more than 1 or add the master=true/py-autoreload=true option, then only one process will store the model, when the request is executed by any other uwsgi worker, harakiri will fire on timeout.
Question: How to make the model initialization occurs when each UWSGI worker spawns separately? (I attach the current ini file with the config for one worker and one thread on which everything works, I attach below.)

[uwsgi]
socket = 0.0.0.0:8000
protocol = http
chdir = /opt/project
module = wsgi:application
threads =  1
workers = 1
listen = 128
harakiri = 60
single-interpreter = true

Answer 1

Found the answer to the question.
If someone encounters such a problem in the future, then you can solve it using the lazy-apps=true option.
In this case, the application is initialized in each worker.
At the moment, the ini file looks like:

[uwsgi]
socket = 0.0.0.0:8000
protocol = http
chdir = /opt/project
module = wsgi:application
cheaper = 2
processes = 16
listen = 128
harakiri = 60
master = true
reaper = true
enable-threads = true
single-interpreter = true
py-autoreload = 1
lazy-apps = true

Answer 2

lazy-apps = true
There was a problem with starting threads in each worker. This mode solved the problem.