ml-config-manager

ML Configuration Management

Generate and manage Hydra configuration files for machine learning experiments with PyTorch Lightning.

Quick Reference

Template Files Available:

• templates/model-config.yaml - Model architecture configuration
• templates/data-config.yaml - Dataset and DataLoader configuration
• templates/trainer-config.yaml - PyTorch Lightning Trainer configuration
• templates/experiment-config.yaml - Complete experiment composition
• templates/gnn-config.yaml - PyTorch Geometric GNN configuration
• templates/sweep-config.yaml - Hyperparameter sweep with Optuna

Configuration Types

1. Model Configuration

Location: configs/model/<name>.yaml

What to ask:

• Model architecture (CNN, Transformer, GNN, MLP, etc.)
• Input/output dimensions
• Hidden dimensions and layers
• Activation functions
• Normalization layers (batch norm, layer norm)
• Dropout rates
• Optimizer type and parameters
• Learning rate scheduler

Generate from template:

_target_: src.models.<name>.Model

# Architecture
input_dim: 784
hidden_dims: [512, 256, 128]
output_dim: 10
activation: relu
dropout: 0.2
batch_norm: true

# Optimizer
optimizer:
  _target_: torch.optim.AdamW
  lr: 0.001
  weight_decay: 0.0001
  betas: [0.9, 0.999]

# Scheduler
scheduler:
  _target_: torch.optim.lr_scheduler.CosineAnnealingLR
  T_max: 100
  eta_min: 1e-6

Common optimizers:

• torch.optim.AdamW - Default choice, good for most tasks
• torch.optim.Adam - Classic optimizer
• torch.optim.SGD - With momentum for vision tasks
• torch.optim.RMSprop - Good for recurrent networks

Common schedulers:

• CosineAnnealingLR - Smooth cosine decay
• ReduceLROnPlateau - Reduce on metric plateau
• OneCycleLR - Super-convergence with 1cycle policy
• StepLR - Step decay at fixed intervals
• ExponentialLR - Exponential decay

See templates/model-config.yaml for complete template.

2. Data Configuration

Location: configs/data/<name>.yaml

What to ask:

• Dataset name/type
• Batch size
• Number of workers
• Data augmentation strategy
• Train/val/test split ratios
• Preprocessing requirements

Generate from template:

_target_: src.data.<name>.DataModule

# Dataset
dataset_name: "mnist"
data_dir: "data/"
download: true

# DataLoader
batch_size: 128
num_workers: 4
pin_memory: true
persistent_workers: true
prefetch_factor: 2

# Splits
train_val_test_split: [0.8, 0.1, 0.1]
shuffle_train: true

# Augmentation
augmentation:
  random_crop: true
  horizontal_flip: true
  normalize: true
  mean: [0.485, 0.456, 0.406]
  std: [0.229, 0.224, 0.225]

See templates/data-config.yaml for complete template with all augmentation options.

3. Trainer Configuration

Location: configs/trainer/<name>.yaml

What to ask:

• Max epochs
• Precision (32, 16-mixed, bf16-mixed)
• Accelerator (auto, gpu, cpu, mps)
• Number of devices (GPUs)
• Strategy (auto, ddp, fsdp, deepspeed)
• Gradient clipping value
• Accumulation steps
• Validation frequency

Generate from template:

_target_: pytorch_lightning.Trainer

# Training duration
max_epochs: 100
min_epochs: 1

# Hardware
accelerator: gpu
devices: 1
precision: 16-mixed

# Distributed
strategy: auto

# Optimization
gradient_clip_val: 1.0
accumulate_grad_batches: 1

# Validation
check_val_every_n_epoch: 1
num_sanity_val_steps: 2

# Callbacks
callbacks:
  - _target_: pytorch_lightning.callbacks.ModelCheckpoint
    monitor: "val/loss"
    mode: "min"
    save_top_k: 3
    save_last: true

  - _target_: pytorch_lightning.callbacks.EarlyStopping
    monitor: "val/loss"
    patience: 10
    mode: "min"

  - _target_: pytorch_lightning.callbacks.LearningRateMonitor
    logging_interval: "step"

  - _target_: pytorch_lightning.callbacks.RichProgressBar

See templates/trainer-config.yaml for complete template with all callbacks and options.

4. Logger Configuration

Location: configs/logger/<name>.yaml

W&B (Recommended):

_target_: pytorch_lightning.loggers.WandbLogger

project: "my-ml-project"
name: null  # Auto-generated from config
save_dir: "logs/"
log_model: true
save_code: true

TensorBoard:

_target_: pytorch_lightning.loggers.TensorBoardLogger

save_dir: "logs/"
name: null
version: null
log_graph: true
default_hp_metric: false

MLflow:

_target_: pytorch_lightning.loggers.MLFlowLogger

experiment_name: "my-ml-project"
tracking_uri: "file:./mlruns"
log_model: true

5. Experiment Configuration

Location: configs/experiment/<name>.yaml

Compose complete experiments from existing configs:

# @package _global_

# Experiment metadata
name: "resnet50_imagenet_baseline"
description: "Baseline ResNet-50 on ImageNet with standard augmentation"
tags: ["baseline", "resnet", "imagenet"]

# Compose from existing configs
defaults:
  - override /model: resnet50
  - override /data: imagenet
  - override /trainer: gpu_ddp
  - override /logger: wandb

# Seed for reproducibility
seed: 42

# Model overrides (specific to this experiment)
model:
  hidden_dims: [2048, 1024, 512]
  dropout: 0.3
  optimizer:
    lr: 0.001

# Data overrides
data:
  batch_size: 256
  num_workers: 8

# Trainer overrides
trainer:
  max_epochs: 200
  devices: 4

# Logger configuration
logger:
  wandb:
    project: "imagenet-experiments"
    tags: ${tags}
    notes: ${description}

See templates/experiment-config.yaml for complete template.

6. Hyperparameter Sweep Configuration

Location: configs/sweep/<name>.yaml

What to ask:

• Parameters to sweep (model, data, training params)
• Search strategy (grid, random, bayesian/Optuna)
• Search space (ranges, choices)
• Optimization metric and direction
• Number of trials

Generate from template (Optuna Bayesian Optimization):

defaults:
  - override hydra/sweeper: optuna

hydra:
  sweeper:
    _target_: hydra_plugins.hydra_optuna_sweeper.optuna_sweeper.OptunaSweeper

    direction: minimize  # minimize or maximize
    n_trials: 50
    n_jobs: 1

    study_name: "mlp_optimization"
    storage: null  # null for in-memory, or "sqlite:///optuna.db"

    # Sampler
    sampler:
      _target_: optuna.samplers.TPESampler
      seed: 42
      n_startup_trials: 10

    # Search space
    params:
      model.hidden_dims:
        type: categorical
        choices:
          - [512, 256]
          - [1024, 512, 256]
          - [2048, 1024, 512]

      model.dropout:
        type: float
        low: 0.0
        high: 0.5
        step: 0.05

      model.optimizer.lr:
        type: float
        low: 0.0001
        high: 0.01
        log: true

      data.batch_size:
        type: categorical
        choices: [64, 128, 256]

# Metric to optimize
optimized_metric: "val/loss"

Run sweep:

python src/train.py --multirun --config-name sweep/<name>

Alternative: Grid search (no Optuna needed):

python src/train.py --multirun \
  model.hidden_dims="[512,256],[1024,512,256]" \
  model.dropout=0.1,0.2,0.3,0.4 \
  model.optimizer.lr=0.001,0.01,0.1

See templates/sweep-config.yaml for complete template with all options.

7. PyTorch Geometric Specific Configs

Location: configs/model/gnn/<name>.yaml

For Graph Neural Networks with PyTorch Geometric:

_target_: src.models.gnn.GNNModel

# GNN architecture
conv_type: GCNConv  # GCNConv, GATConv, SAGEConv, GINConv, TransformerConv
num_layers: 3
hidden_channels: 128
out_channels: 64

# GNN-specific
aggr: "add"  # add, mean, max
normalize: true
dropout: 0.2
jk_mode: null  # null, cat, max, lstm

# Attention (for GAT/TransformerConv)
heads: 8
concat_heads: true

# Global pooling (for graph-level tasks)
global_pool: "mean"  # mean, max, add, attention, set2set

# Task configuration
task: "node_classification"  # node_classification, graph_classification, link_prediction
num_classes: 7

# Optimizer
optimizer:
  _target_: torch.optim.AdamW
  lr: 0.01
  weight_decay: 5e-4  # Higher weight decay for GNNs

Corresponding graph data config:

_target_: src.data.graph_datamodule.GraphDataModule

dataset_name: "Cora"
data_dir: "data/graphs/"
batch_size: 32

# Sampling (for large graphs)
use_sampling: false
num_neighbors: [15, 10, 5]

See templates/gnn-config.yaml for complete template.

Configuration Best Practices

1. Naming Conventions

• Use descriptive names: resnet50_pretrained.yaml, not model1.yaml
• Hierarchical naming: trainer/gpu_single.yaml, trainer/gpu_multi.yaml
• Domain prefixes: data/vision/, data/nlp/, data/graph/

2. Modularity

• Create reusable components
• Use defaults composition for experiments
• Override only what's necessary
• Keep configs DRY (Don't Repeat Yourself)

3. Documentation

Add comments explaining non-obvious parameters:

# Use higher weight decay for GNNs to prevent overfitting
weight_decay: 5e-4

# TPESampler needs startup trials for warm-up
n_startup_trials: 10

4. Type Safety

Use Pydantic models for validation:

from pydantic import BaseModel

class ModelConfig(BaseModel):
    input_dim: int
    hidden_dims: list[int]
    output_dim: int
    dropout: float

5. Versioning

• Track all config changes in git
• Tag configs with experiment versions
• Use meaningful commit messages for config changes

6. Sensible Defaults

• Provide defaults for all optional parameters
• Use industry-standard hyperparameters as defaults
• Document why defaults were chosen

Validation Checklist

After generating configs, validate:

• YAML syntax is valid (no tabs, correct indentation)
• All _target_ paths exist and are importable
• No circular dependencies in defaults
• Config loads without errors: python src/train.py --cfg job
• Print resolved config: python src/train.py --cfg job --print-config
• All file paths are correct (data_dir, save_dir, etc.)
• Hyperparameter ranges are reasonable
• Batch size fits in GPU memory

Common CLI Overrides

# Override single parameter
python src/train.py model.lr=0.01

# Override multiple parameters
python src/train.py model.lr=0.01 data.batch_size=256

# Use specific experiment config
python src/train.py experiment=resnet50_imagenet

# Override nested parameters
python src/train.py model.optimizer.weight_decay=1e-4

# Use different config group
python src/train.py model=transformer data=wikitext

# Multirun (grid search)
python src/train.py --multirun model.dropout=0.1,0.2,0.3

# Print resolved config
python src/train.py --cfg job

# Print config with overrides
python src/train.py model.lr=0.01 --cfg job

Debugging Config Issues

Config doesn't load:

# Check YAML syntax
python -c "import yaml; yaml.safe_load(open('configs/model/mymodel.yaml'))"

# Validate with Hydra
python src/train.py --cfg job --config-name myconfig

Import errors:

• Verify _target_ paths are correct
• Check all modules are importable
• Use absolute imports: src.models.mlp.MLP not models.mlp.MLP

Defaults not resolving:

• Check defaults order (later overrides earlier)
• Use override keyword for conflicting groups
• Verify group names match directory structure

Variable interpolation not working:

• Use ${var} syntax for interpolation
• Check variable exists in config
• Use ${oc.env:VAR} for environment variables

Integration with Training

Once configs are created, use them in training:

import hydra
from omegaconf import DictConfig

@hydra.main(version_base=None, config_path="configs", config_name="config")
def train(cfg: DictConfig):
    # Instantiate components from config
    model = hydra.utils.instantiate(cfg.model)
    datamodule = hydra.utils.instantiate(cfg.data)
    trainer = hydra.utils.instantiate(cfg.trainer)

    # Train
    trainer.fit(model, datamodule=datamodule)

Success Criteria

• Configuration files created in correct locations
• YAML syntax is valid
• All _target_ paths resolve correctly
• Config loads without errors
• Overrides work as expected
• Documentation updated (if needed)
• Tested with actual training run

Configuration is ready for experimentation!