Implementation:Facebookresearch Audiocraft AudioCraftEnvironment

Overview

AudioCraftEnvironment is the singleton class that encapsulates all cluster-level and team-level configuration for the AudioCraft training infrastructure. It resolves filesystem paths for Dora experiment tracking, reference files, and dataset path mappings across different compute environments.

Source Location

Property	Value
Source file	`audiocraft/environment.py` lines 25-177
Import	`from audiocraft.environment import AudioCraftEnvironment`
Module	`audiocraft.environment`

API

Singleton Access

AudioCraftEnvironment.instance() -> AudioCraftEnvironment

Returns the singleton instance. Creates it on first call by:

Reading AUDIOCRAFT_TEAM env var (default: "default")
Auto-detecting cluster type via _guess_cluster_type()
Loading team YAML config from config/teams/{team}.yaml (or path from AUDIOCRAFT_CONFIG)
Compiling dataset mapper regex rules from the cluster config

Classmethods

AudioCraftEnvironment.get_team() -> str
AudioCraftEnvironment.get_cluster() -> str
AudioCraftEnvironment.get_dora_dir() -> Path
AudioCraftEnvironment.get_reference_dir() -> Path
AudioCraftEnvironment.get_slurm_exclude() -> Optional[str]
AudioCraftEnvironment.get_slurm_partitions(partition_types: Optional[List[str]] = None) -> str
AudioCraftEnvironment.resolve_reference_path(path: Union[str, Path]) -> Path
AudioCraftEnvironment.apply_dataset_mappers(path: str) -> str
AudioCraftEnvironment.reset() -> None

Environment Variables

Variable	Purpose	Default
`AUDIOCRAFT_TEAM`	Selects the team configuration	`"default"`
`AUDIOCRAFT_CLUSTER`	Overrides auto-detected cluster type	Auto-detected
`AUDIOCRAFT_CONFIG`	Path to custom team YAML config	`config/teams/{team}.yaml`
`AUDIOCRAFT_DORA_DIR`	Overrides Dora experiment directory	From team YAML config
`AUDIOCRAFT_REFERENCE_DIR`	Overrides reference file directory	From team YAML config

Configuration Files

config/teams/default.yaml

default:
  dora_dir: /tmp/audiocraft_${oc.env:USER}
  partitions:
    global: debug
    team: debug
  reference_dir: /tmp
darwin:
  dora_dir: /tmp/audiocraft_${oc.env:USER}
  partitions:
    global: debug
    team: debug
  reference_dir: /tmp

Team-specific configs (e.g., config/teams/labs.yaml) may define additional cluster entries with custom paths, partition names, and dataset mappers.

Inputs and Outputs

Inputs:

Environment variables (AUDIOCRAFT_TEAM, AUDIOCRAFT_DORA_DIR, AUDIOCRAFT_REFERENCE_DIR, AUDIOCRAFT_CLUSTER, AUDIOCRAFT_CONFIG)
Team YAML configuration files (e.g., config/teams/default.yaml)
Runtime cluster detection from audiocraft.utils.cluster._guess_cluster_type()

Outputs:

Singleton instance providing:
- dora_dir -- resolved Path for Dora experiment storage
- reference_dir -- resolved Path for shared reference files
- cluster -- string identifier for the detected cluster
- team -- string identifier for the configured team
- Dataset path mapping via regex substitution rules

Usage in Training Pipeline

The environment is initialized very early in the training pipeline. In audiocraft/train.py:

from .environment import AudioCraftEnvironment

# ... later at module level:
main.dora.dir = AudioCraftEnvironment.get_dora_dir()

The resolve_reference_path method is used throughout the codebase to resolve //reference prefixed paths:

# In audiocraft/metrics/fad.py
self.model_path = AudioCraftEnvironment.resolve_reference_path(model_path)

# In audiocraft/utils/checkpoint.py
path = AudioCraftEnvironment.resolve_reference_path(path)

Dependencies

omegaconf -- for loading and parsing YAML configuration
audiocraft.utils.cluster._guess_cluster_type -- for automatic cluster detection

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment