Docker deployment -- No valid model found at /app/models!

Rasa Open Source
1

First of all, I’m brand new to Rasa…and development in general. :slight_smile:

I’m trying to deploy my Rasa (OS) assistant as a multi-container web app running on a VM. I’ve configured an NLU server, an action server, and a web server (using Nginx). While the action and web servers appear to be running normally, my NLU server can’t seem to locate the trained model.

Here’s the relevant part of my docker-compose.yml file:

version: '3'

services:

  nlu_server:
    container_name: "nlu_server"
    build:
      context: backend
    volumes:
      - ./backend:/app
      - ./backend/data:/app/data
      - ./backend/models:/app/models
    ports:
      - "5005:5005"

  action_server:
    ...

  nginx:
    ...

My backend directory structure looks like this:

|-- backend
|   |-- data
|   |   |-- nlu.yml
|   |   |-- rules.yml
|   |   |-- stories.yml
|   |-- models
|   |   |-- model_0.tar.gz
|   |-- config.yml
|   |-- credentials.yml
|   |-- domain.yml
|   |-- endpoints.yml
|   |-- Dockerfile

And here’s my Dockerfile for the backend context:

FROM rasa/rasa:3.3.1

WORKDIR /app

USER root

COPY ./data /app/data
COPY ./models /app/models
COPY . /app

RUN rasa train --force --out /app/models

ENTRYPOINT ["rasa", "run", "-m", "/app/models", "--enable-api", "--cors", "*", "--debug"]

I can confirm that running docker compose build does, indeed, train a new Rasa model (call it model_1.tar.gz) and (claims to) save it to /app/models (which already contains model_0.tar.gz since it’s copied from the context). However, when I run docker compose up, the last few lines of the log are:

nlu_server | 2023-04-07 21:08:08 INFO     rasa.core.processor  - Loading model /app/models/model_0.tar.gz...
nlu_server | /opt/venv/lib/python3.8/site-packages/rasa/shared/utils/io.py:98: UserWarning: No valid model found at /app/models!
nlu_server | 2023-04-07 21:08:08 INFO     root  - Rasa server is up and running.

So Rasa is attempting to load the pre-existing model (the one included with the build context), and finds it to be invalid for some reason…? (To be honest, this is why I included a rasa train command in the Dockerfile in the first place…)

EDIT: After running docker exec -it <container id> bash, I can see that the model trained during the build is not in the container.

What I am doing wrong?

2

Do you have found a solution for this? I’m having the same issue.

3

Unfortunately, I did not. To bypass the issue, I ended up including a rasa train statement in the Dockerfile.