Hydra Experiment Config

When to Use

• Setting up Hydra config structure for a new ML project
• Adding new experiment variants (model sizes, training schedules, datasets)
• Debugging config override resolution or composition errors
• Reviewing whether runtime constants are properly externalized to config

Core Principles

Config is King

Every runtime constant lives in YAML config, never hardcoded in Python. Checkpoint metadata is diagnostic only — the config file is the single source of truth for reproducing a run.

Hierarchical Groups with Flat Aliases

Organize configs into semantic groups, but provide flat backward-compatible aliases so users don't need to memorize the full path:

# conf/training/default.yaml
run:
  iter_num: 100000
  seed: 42
optim:
  lr: 1e-4
  weight_decay: 0.01
loader:
  batch_size: 16
  num_workers: 4

# Flat aliases (resolved at runtime via OmegaConf interpolation or code)
iter_num: ${run.iter_num}
lr: ${optim.lr}
batch_size: ${loader.batch_size}

Both of these work:

python train.py training.loader.batch_size=8   # hierarchical
python train.py training.batch_size=8           # flat alias

One Entry Point, Mode Switch

Use a single entry point with a mode field to dispatch:

@hydra.main(config_path="conf", config_name="config")
def main(cfg):
    if cfg.mode == "train":
        train(cfg)
    elif cfg.mode == "evaluate":
        evaluate(cfg)
    elif cfg.mode == "recognize":
        recognize(cfg)

Patterns

Config Group Layout

conf/
  config.yaml               # Top-level: defaults list + mode
  model/
    base.yaml                # Default model
    tiny.yaml                # Small model for fast iteration
    dryrun.yaml              # Minimal model for smoke tests
  training/
    default.yaml             # Standard training config
    dryrun.yaml              # Smoke test (few iterations)
    fastrun.yaml             # Quick validation (frequent eval)
    fullrun.yaml             # Production run
  data/
    dataset_a.yaml           # Dataset-specific loader config
    dataset_b.yaml

Defaults List

# conf/config.yaml
defaults:
  - model: base
  - training: default
  - data: dataset_a
  - _self_

mode: train

Derived Values

Compute derived values from config, never from magic numbers:

# Good: derived from config
fps = cfg.model.encoder.sample_rate / cfg.model.encoder.hop_samples

# Bad: magic number
fps = 25

Eliminating Module-Level Constants

Module-level constants that duplicate config values must be replaced with config-driven parameters:

# Bad: hardcoded constant duplicates config
TARGET_SAMPLE_RATE = 24000  # lives in model.encoder.sample_rate

def resample(waveform, orig_sr):
    if orig_sr != TARGET_SAMPLE_RATE:
        ...

# Good: accept from config, use function default only as safety net
def resample(waveform, orig_sr, target_sr: int):
    if orig_sr != target_sr:
        ...

Function parameter defaults like target_sr: int = 24000 are acceptable as fallbacks, but the actual value must always be threaded from config at call sites. Use functools.partial to bind config values into pipeline callbacks that can't accept extra arguments directly.

CLI Override Patterns

# Single override
python train.py training.optim.lr=3e-4

# Switch config group
python train.py model=tiny training=fastrun

# Multi-run sweep
python train.py --multirun training.optim.lr=1e-4,3e-4,1e-3

Anti-Patterns

• Hardcoded constants in Python: If a value might change between experiments, it belongs in config.
• Sbatch scripts with inline hyperparameters: Hyperparameters live in conf/ YAML; sbatch scripts only set environment variables and call python train.py with config group overrides.
• Checkpoint as config source: Checkpoints may store config snapshots for diagnostics, but the YAML files are the authority for reproduction. Never load training config from a checkpoint to resume — use the original config file.
• Deeply nested overrides without aliases: If users frequently override training.optim.scheduler.warmup_steps, provide a flat alias training.warmup_steps.
• Mixing config and argparse: Choose one. Hydra replaces argparse entirely.

See Also

• fail-fast-ml-engineering — Config as single source of truth, preflight validation
• wandb-experiment-tracking — Logging resolved Hydra config to W&B at init