Status: Needs Review

This page has not been reviewed for accuracy and completeness. Content may be outdated or contain errors.

Hydra Inheritance & Overrides¶

Config inheritance, variable interpolation, and override mechanisms.

Config Inheritance¶

Simple Inheritance¶

Base config: configs/training/base.yaml

seed: 42

trainer:
  max_epochs: 5
  accelerator: auto
  devices: 1

optimizer:
  name: adamw
  betas: [0.9, 0.999]

Variant: configs/training/high_lr.yaml

defaults:
  - base
  - _self_

optimizer:
  lr: 0.001           # Add new field
  weight_decay: 0.01  # Add new field

Result:

seed: 42                        # From base
trainer:
  max_epochs: 5                 # From base
  accelerator: auto             # From base
  devices: 1                    # From base
optimizer:
  name: adamw                   # From base
  betas: [0.9, 0.999]          # From base
  lr: 0.001                     # From high_lr
  weight_decay: 0.01            # From high_lr

Multi-Level Inheritance¶

# Level 1: configs/training/base_optimizer.yaml
optimizer:
  name: adamw
  betas: [0.9, 0.999]

# Level 2: configs/training/base_training.yaml
defaults:
  - base_optimizer
  - _self_
seed: 42
trainer:
  max_epochs: 5

# Level 3: configs/training/custom_training.yaml
defaults:
  - base_training
  - _self_
optimizer:
  lr: 0.001
trainer:
  max_epochs: 100  # Override base

Override Behavior¶

Hydra uses merge strategy by default — only explicitly set fields are overridden:

# Base
data:
  train_ids: [0, 2, 3]
  val_ids: [1, 5]
  batch_size: 2

# Override (only batch_size changes)
defaults:
  - base
  - _self_
data:
  batch_size: 16  # train_ids and val_ids unchanged

Variable Interpolation¶

Simple Interpolation¶

name: my_experiment
output_dir: ./outputs/${name}
# Resolves to: ./outputs/my_experiment

training:
  trainer:
    default_root_dir: ${output_dir}

Environment Variables¶

data:
  cu3s_file_path: ${oc.env:DATA_ROOT}/Lentils_000.cu3s

  # With fallback:
  cu3s_file_path: ${oc.env:DATA_ROOT,./data/Lentils}/Lentils_000.cu3s

OmegaConf Resolvers¶

Resolver	Purpose	Example
`oc.env`	Environment variable	`${oc.env:DATA_ROOT}`
`oc.decode`	Python expression	`${oc.decode:"1 if x else 0"}`
`oc.create`	Create object	`${oc.create:datetime.datetime.now}`

Cross-Group References¶

defaults:
  - /pipeline/anomaly/rx@pipeline: rx_statistical
  - /training@training: default
  - _self_

output_dir: ./outputs/${pipeline.metadata.name}
experiment_seed: ${training.seed}

Override Mechanisms¶

1. Config-Level Overrides¶

defaults:
  - /pipeline/anomaly/rx@pipeline: rx_statistical
  - /data@data: lentils
  - /training@training: default
  - _self_  # ← Must be last

data:
  train_ids: [0, 1, 2]
  batch_size: 16
training:
  optimizer:
    lr: 0.0001

2. Command-Line Overrides¶

python train.py training.optimizer.lr=0.001

python train.py \
    training.trainer.max_epochs=100 \
    training.optimizer.lr=0.001 \
    data.batch_size=16

# List/dict assignment
python train.py data.train_ids=[0,1,2,3]
python train.py training.scheduler={name:reduce_on_plateau,patience:10}

3. Config Group Selection¶

python train.py pipeline=channel_selector
python train.py data=custom_dataset
python train.py training=high_lr

4. Programmatic Overrides¶

from omegaconf import OmegaConf

@hydra.main(config_path="../configs", config_name="trainrun/default_gradient")
def main(cfg: DictConfig):
    cfg.training.optimizer.lr = 0.0001

    overrides = OmegaConf.create({
        "training": {"optimizer": {"lr": 0.0001}, "trainer": {"max_epochs": 100}}
    })
    cfg = OmegaConf.merge(cfg, overrides)

    config_dict = OmegaConf.to_container(cfg, resolve=True)