dota-draft

ML Application for drafting heroes in the Dota 2 game.

Data Source

OpenDota API is scraped to retrieve match ids along with the teams' drafts and outcome.

Technologies used

• Google Cloud Function
• Google Cloud Pub/Sub
• Google Cloud Scheduler
• Google Cloud Dataflow / Apache Beam
• Google Cloud AI Platform
• Weights and Biases

Workflow Summary

1. Trigger data collection from CloudScheduler using a Pub/Sub topic
   • Cloud function collects matches from the OpenDota API and saves them to CloudStorage
2. Trigger manually a batch processing of data that filters matches of interest and saves them to CloudStorage
   • These files are set up for a training job and are split into train/val/test folders
3. Train a model to predict the outcome of a match given the draft
   • Currently saving the model using Weights & Biases, working on training using Cloud AI Platform
4. The model is stored online and can be retrieved for evaluation

Pending

5. Use the trained model to suggest heroes in a web UI using torch-js

Configs

In order for the project to work as intended, you'll need to set up the following files:

• draft/configs/gar.yaml
• draft/configs/gcs.yaml
• draft/configs/opendota.yaml
• draft/configs/wandb.yaml

See the specific README for more details.