neusky/docker-compose.yml at main · JADGardner/neusky · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
# docker-compose.yml — NeuSky
#
# GPU-accelerated research container for NeuSky outdoor scene reconstruction.
#
# Quick reference:
#   Build the image:        docker compose build research
#   Start a research shell: docker compose run research bash
#   Train NeuSky:           docker compose run research ns-train neusky --data /workspace/data/<scene>
#
# Environment variables (set in shell or .env file):
#   DATA_PATH          — path to datasets (default: ./data)
#   OUTPUTS_PATH       — path to training outputs (default: ./outputs)
#   MODEL_STORAGE_PATH — path to pretrained models (default: ./model-storage)

# Two-level container design: this standalone compose mounts the neusky repo
# at /workspace so it is clone-alone-runnable (project name `neusky`). The phd
# umbrella compose is separate (project name `phd`), mounting the whole tree at
# /workspace/phd. The explicit name below keeps the two projects from colliding
# and makes wrong-cwd invocations visible (containers are named neusky-*).
name: neusky

services:
  research:
    build:
      context: .
      dockerfile: docker/Dockerfile
    volumes:
      # Mount the project root
      - .:/workspace
      # Mount external data, models, and outputs (override via .env or env vars)
      - ${DATA_PATH:-./data}:/workspace/data
      - ${MODEL_STORAGE_PATH:-./model-storage}:/workspace/model-storage
      - ${OUTPUTS_PATH:-./outputs}:/workspace/outputs
      # Mount entrypoint from host (avoids image rebuild for entrypoint changes)
      - ./docker/entrypoint.sh:/entrypoint.sh:ro
      # Mount wandb credentials
      - ~/.netrc:/tmp/.netrc:ro
      - ~/.cache/wandb:/tmp/home/.cache/wandb
    user: "${UID:-1000}:${GID:-1000}"
    working_dir: /workspace
    ports:
      - "7007:7007"
    # Shared memory for PyTorch DataLoader multiprocessing
    ipc: host
    # Enable GPU access via the NVIDIA Container Toolkit
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      - CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-0}
      - PYTHONUNBUFFERED=1
      - PYTHONPATH=/workspace
      - HOME=/tmp/home
      - NETRC=/tmp/.netrc
      - MODEL_STORAGE_PATH=/workspace/model-storage
      - DATA_PATH=/workspace/data
      - OUTPUTS_PATH=/workspace/outputs
      - PROJECT_ROOT=/workspace
      # Legacy checkpoints contain numpy scalars that fail with PyTorch 2.6+ weights_only=True
      - TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1