docker-compose up --build

Check routes

flask --app db_service routes
flask --app model_service/resnet50 routes
flask --app xai_service/pytorch_cam routes

Run debug mode

flask --app model_service/resnet50 --debug run -p 5001
flask --app db_service --debug run -p 5002
flask --app xai_service/pytorch_cam --debug run -p 5003
flask --app evaluation_service --debug run -p 5004

XAI Service Frontend

This is the frontend for the eXplainable AI service.

The project uses Next.js framework, styled with Tailwindcss, and Prisma ORM.

It is hosted on Vercel.

Development Prequisites:

• Node >= 18.x
• npm >= 8.18.0
• Docker Engine

Quickstart

• Configure your .env environment variables from .env.template
• Clone the project: git clone https://github.com/ZeruiW/XAI-Service
• Change directory into cloned folder: cd XAI-Service
• Install node depencies: npm i
• Start development server: npm run dev

XAI Service Backend

Example XAI task setting

{"method_name":"hirescam","data_set_name":"azure_cog","data_set_group_name":"1500","model_name":"resnet50"}

Example evaluation task setting

{"explanation_task_ticket":"CRx1G9HCKKsh4Qs.9OF7NXAJSL"}

Local Dev

MongoDB

Please have the mongo.dev.conf or mongo.pred.conf under the backend folder.

conn_str=<<your mongodb url str>>

Azure Blob Service

Please have the az_blob_connection_str.json under the backend folder:

1. backend/central/central_storage/tmp/az_blob_connection_str.json;
2. backend/db_service/azure_blob/azure_blob_storage/tmp/az_blob_connection_str.json

The file is private, if you need them, please contact JUN. Or you can deploy your own Azure Blob Service.

Run In Different Env Mode

Please use:

flask --app 'backend/central:create_app("dev")' run -p 5006

or

flask --app 'backend/central:create_app("pred")' run -p 5006

to start the flask application.

Run Docker in Dev Env

1. Start-Up Local MySQL

If you are the first time, please also create a volume for MySQL.

docker volume create xaifw-mysql

Then:

docker compose -f backend/db_service/docker-compos-mysql.yaml up -d

2. Volume for All the Services

docker volume create xai_fw_volumes

3. Bring Up Services

docker compose -f backend/docker-compose.yml -f backend/docker-compose-dev.yml up --build

Or for single service:

docker compose -f backend/docker-compose.yml -f backend/docker-compose-dev.yml up [service_name] --build

Critical Update

Jan 20th

1. Azure Blob Service;

2. Densenet121 service:

   flask --app backend/model_service/densenet121 run -p 5010

3. multiple cam method support

   To start grad-cam:

   flask --app 'backend/xai_service/pytorch_cam:create_app(cam_method="grad-cam")' run -p 5003

   or just

   flask --app backend/xai_service/pytorch_cam run -p 5003

   Then the service endpoint is same as before:

   http://127.0.0.1:5003/xai/pt_cam

   To start other cams, like grad-camew:

   flask --app 'backend/xai_service/pytorch_cam:create_app(cam_method="grad-camew")' run -p 5011

   The service endpoint will be:

   http://127.0.0.1:5011/xai/pt_cam/grad-camew

   Cam method List:

   if cam_method == None or cam_method == 'grad-cam':
       cam = GradCAM(**cam_kws)
   
   if cam_method == 'hirescam':
       cam = HiResCAM(**cam_kws)
   
   if cam_method == 'scorecam':
       cam = ScoreCAM(**cam_kws)
   
   if cam_method == 'grad-campp':
       cam = GradCAMPlusPlus(**cam_kws)
   
   if cam_method == 'ablationcam':
       cam = AblationCAM(**cam_kws)
   
   if cam_method == 'xgrad-cam':
       cam = XGradCAM(**cam_kws)
   
   if cam_method == 'eigencam':
       cam = EigenCAM(**cam_kws)
   
   if cam_method == 'eigengrad-cam':
       cam = EigenGradCAM(**cam_kws)
   
   if cam_method == 'layercam':
       cam = LayerCAM(**cam_kws)
   
   if cam_method == 'fullgrad':
       cam = FullGrad(**cam_kws)
   
   if cam_method == 'grad-camew':
       cam = GradCAMElementWise(**cam_kws)

​

Requirements

Platforms

• Linux x86-64:
  • Ubuntu 22.04 LTS (Best)
  • Debian 11 "Bullseye" Stable branch
  • RHEL 8
  • Arch
• macOS x86-64 / ARM64 (local dev only)
• Windows x86-64:
  • Windows 10 >= 1909 update
  • Windows 11

Hardware Acceleration

• NVIDIA CUDA Library
• Vulkan

TODOs

Tasks and child tasks are priority tagged starting from 0 as the highest priority. E.g. P0, P1,...Pn. Finished tasks are stripped of the priority tag.

For maintainers, it is advised to follow the Notion documentation (shared internally) as the single source of truth.

Progress moved to Github Projects