docker-compose.yml

# ExPhil Docker Compose Configuration
# Usage:
#   docker-compose build                      # Build the image
#   docker-compose run train-production       # Production preset (best quality)
#   docker-compose run train-standard         # Standard preset (balanced)
#   docker-compose run train-mewtwo           # Mewtwo specialist
#   docker-compose run shell                  # Interactive Elixir shell

services:
  # =============================================================================
  # GPU Training with Presets
  # =============================================================================

  # Production preset - maximum quality training
  # ~50 epochs, Mamba backbone, all features enabled
  train-production:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
      - WANDB_API_KEY=${WANDB_API_KEY:-}
    command: >
      mix run scripts/train_from_replays.exs
      --preset production
      --replays /app/replays
      --checkpoint /app/checkpoints

  # Production with online robustness (for Slippi netplay)
  # Adds frame delay augmentation (0-18 frames)
  train-production-online:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
      - WANDB_API_KEY=${WANDB_API_KEY:-}
    command: >
      mix run scripts/train_from_replays.exs
      --preset production
      --online-robust
      --replays /app/replays
      --checkpoint /app/checkpoints

  # GPU standard preset - good balance of speed and quality
  # ~20 epochs, all augmentation features
  train-standard:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset gpu_standard
      --replays /app/replays
      --checkpoint /app/checkpoints

  # GPU quick - fast validation that GPU training works
  # ~3 epochs, 20 files
  train-quick:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset gpu_quick
      --replays /app/replays
      --checkpoint /app/checkpoints

  # =============================================================================
  # Character-Specific Training
  # =============================================================================

  # Mewtwo specialist - 90-frame window for teleport recovery
  train-mewtwo:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset mewtwo
      --character MEWTWO
      --online-robust
      --replays /app/replays
      --checkpoint /app/checkpoints

  # Ganondorf specialist - 60-frame window for spacing
  train-ganondorf:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset ganondorf
      --character GANONDORF
      --online-robust
      --replays /app/replays
      --checkpoint /app/checkpoints

  # Link specialist - 75-frame window for projectiles
  train-link:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset link
      --character LINK
      --online-robust
      --replays /app/replays
      --checkpoint /app/checkpoints

  # Game & Watch specialist - 45-frame window (L-cancels: fair, dair)
  train-gameandwatch:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset gameandwatch
      --character GAME_AND_WATCH
      --online-robust
      --replays /app/replays
      --checkpoint /app/checkpoints

  # Zelda specialist - 60-frame window for transform
  train-zelda:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/train_from_replays.exs
      --preset zelda
      --character ZELDA
      --online-robust
      --replays /app/replays
      --checkpoint /app/checkpoints

  # =============================================================================
  # Utilities
  # =============================================================================

  # Replay scanner - analyze character distribution in replay collection
  scan-replays:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    volumes:
      - ./replays:/app/replays:ro
    environment:
      - MIX_ENV=prod
    command: >
      mix run scripts/scan_replays.exs
      --replays /app/replays

  # Learning rate finder
  find-lr:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    command: >
      mix run scripts/find_lr.exs
      --replays /app/replays

  # Interactive Elixir shell for debugging
  shell:
    build:
      context: .
      dockerfile: Dockerfile.gpu
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
      - ./lib:/app/lib:ro
      - ./scripts:/app/scripts:ro
    environment:
      - MIX_ENV=prod
      - CUDA_VISIBLE_DEVICES=0
    stdin_open: true
    tty: true
    command: iex -S mix

  # =============================================================================
  # CPU Testing (Local development without GPU)
  # =============================================================================

  # CPU quick test
  train-cpu-quick:
    build:
      context: .
      dockerfile: Dockerfile.cpu
    volumes:
      - ./replays:/app/replays:ro
      - ./checkpoints:/app/checkpoints:rw
    environment:
      - MIX_ENV=prod
    command: >
      mix run scripts/train_from_replays.exs
      --preset quick
      --replays /app/replays
      --checkpoint /app/checkpoints