1. Brain Piano backends in Modern Architecture

After gaining more knowledge on the cloud infrastructure-based micro-services architecture, I wanted to regain life for the Brain Piano Project. In 2019, I carried two laptops to run the demo - one to run a compiled version of Magenta code and another to run client-side javascript code - with limited knowledge of ML deployment.

Here are criteria of the new service
• The server must be remotely hosted
• The server must be publicly accessible

2. Unexpected Challenge

The most recent Magenta Repo suggest testing the model with the Colab environment due to the specific software(OS, Python environment) dependencies.

However, Colab executes codes in a runtime environment, which is unsuitable for my use case. Moreover, I found that the docker environment, which used to be introduced as an alternative installation method, is moved from their official documentation.

3. Dockerization of Magenta

Thankfully, I came across this amazing repo by xychelsea who recreated Magenta docker-environment. This is my version of docker-compose.yml file to get it work with MacOS(Ventura 13.5.1, M1).

1
2# docker-compose.yml file
3version: "3.7"
4
5networks:
6  default:
7    driver: bridge
8
9services:
10  portainer:
11    image: portainer/portainer-ce:latest
12    ports:
13      - "9000:9000"
14    restart: always
15    volumes:
16      - portainer-data:/data
17      - /var/run/docker.sock:/var/run/docker.sock
18  magenta:
19    platform: linux/amd64
20    image: xychelsea/magenta:latest-jupyter
21    container_name: magenta
22    restart: always
23    ports:
24      - "8888:8888"
25    volumes:
26      - ./volume/magenta:/usr/local/magenta/workspace
27volumes:
28  portainer-data:
29

4. Deploy Code in GCP

I decided to deploy the code to the GCP free instance(e2-micro) to create the publicly accessible server.

To receive the command from the front-end and deliver the message to the dockerized magenta code, I created a simple FlaskAPI server.

The generated midi file from the magenta is then uploaded to Google Drive to share input/output files with the front-end. The diagram below summarizes what is happening inside the GCP instance.

This time, I will not cover the specific steps to deploy the docker container to the GCP instance. Instead, I recommend reading this article to follow through with the details.

5. Result

The picture below shows the public webpage deployed to the GCP instance. This swagger UI webpage shows the HTTP endpoints of Flask API.

When receiving HTTP GET request in /generate endpoint, the backend server generates classical music in the style of Bach with the melody line provided.

However, music generation process is extremely slow - it took around 30 seconds to generate the 10 seconds of music, and it is due to the limited hardware specification of the free GCP instance.(2 vCPU | RAM 1024 MiB | Storage 10Gb Standard (Non-SSD))

I started to search better options to host the ML service, this journey will be covered in the next post.