Config Reference

Config Reference#

This page documents all ChemTorch-specific configuration options that can be set in experiment configs or via the CLI.

If you look for guidance on designing configuration files, see Hydra Config Design which covers ChemTorch-specific extensions to the standard Hydra configuration system.

Note

Defaults are set in conf/base.yaml, however, not all supported options necessarily have a corresponding default value since specifyng key=null in the config is the same as not setting the key at all. Just know that you can still set all options listed below in your experiment config or add them in the CLI via +key.

Execution Control#

tasks

Controls which stages of the pipeline to execute.

  • Type: List of strings

  • Options: fit, test, validate, predict

  • Default: null (must be specified)

tasks:
  - fit
  - test

chemtorch +experiment=graph tasks='[fit,test]'

  • fit: Training and validation

  • test: Evaluate on test set(s)

  • validate: Evaluate on validation set only

  • predict: Run inference without training (see Running Inference).

The predict task runs inference only (no training). Before using it, you must already have a trained model checkpoint — first run fit to produce one, then set load_model: true together with ckpt_path (see Model Loading). For predict the data pipeline must not perform any splitting (set the data splitter to null) and the column mapper must omit the label/target column if it exists. ChemTorch will enforce these automatically for a predict-only run, but declaring them explicitly can make configs clearer and easier to audit.

To persist outputs from a predict-only run, use the keys described under Prediction Saving: predictions_save_path for a single partition (tasks: [predict]) or save_predictions_for plus predictions_save_dir when multiple partitions are available. If instead you want to capture predictions for just one of the train, val, or test partitions during/after training (without a separate predict-only task), configure the same Prediction Saving keys as explained in that section.

Note that to specify a single task you also need to use the list syntax.

tasks:
  - predict

chemtorch +experiment=graph tasks='[predict]'.

seed

Random seed for reproducibility.

  • Type: Integer

  • Default: 0

seed: 42

chemtorch +experiment=graph seed=42

Sets the random seed for Python, NumPy, PyTorch, and PyTorch Lightning to ensure reproducible results.

Logging (Weights & Biases)#

log

Enable or disable Weights & Biases logging.

  • Type: Boolean

  • Default: false

log: true

chemtorch +experiment=graph log=true

project_name

W&B project name for organizing runs.

  • Type: String

  • Default: "chemtorch"

project_name: my_project

chemtorch +experiment=graph project_name=my_project

group_name

W&B group name for organizing related runs (e.g., hyperparameter sweeps).

  • Type: String or null

  • Default: null

group_name: hyperparameter_sweep_1

chemtorch +experiment=graph group_name=my_group

run_name

W&B run name for identifying individual runs.

  • Type: String or null

  • Default: null (W&B generates a random name)

run_name: baseline_experiment

chemtorch +experiment=graph run_name="baseline run"

Model Loading#

load_model

Load a pre-trained model from a checkpoint.

  • Type: Boolean

  • Default: false

  • If true, ckpt_path must be specified.

ckpt_path

Path to the checkpoint file to load.

  • Type: String or null

  • Default: null

  • Required if: load_model=true

load_model: true
ckpt_path: path/to/checkpoint.ckpt

chemtorch +experiment=graph load_model=true ckpt_path=path/to/checkpoint.ckpt

Prediction Saving#

save_predictions_for

Specify which dataset partitions to save predictions for.

  • Type: String, list of strings, or null

  • Options: "train", "val", "test", "predict", "all"

  • Default: null

Only required if mutliple data partations are used. For example:

  • tasks: [fit]: training and validation partitions are available

  • tasks: [fit, test]: training, validation, and test partitions are available

Not required if only a single partition is used (e.g., tasks: [test] or tasks: [predict]).

If set to all, predictions will be saved for all available partitions.

predictions_save_path

Path to save predictions (for single data partition).

  • Type: String or null

  • Default: null

tasks: predict
predictions_save_path: my_experiment/predictions.csv

chemtorch +experiment=graph predictions_save_path=my_experiment/predictions.csv

predictions_save_dir

Directory to save predictions (for multiple data partitions). The predictions will be saved in files named <partition>.csv within this directory, where <partition> is one of train, val, test, or predict.

  • Type: String or null

  • Default: null

  • Requires that save_predictions_for is set.

# Save for multiple partitions
save_predictions_for:
  - train
  - test

chemtorch +experiment=graph save_predictions_for='[train,test]' predictions_save_dir=predictions/

Data Subsampling#

data_module.subsample

Subsample the dataset for quick testing (not a ChemTorch root-level key, but very useful).

  • Type: Float (0.0 to 1.0), int (0 to dataset size) or null

  • Default: null (use full dataset)

# Use 5% of data for quick testing
chemtorch +experiment=graph data_module.subsample=0.05

# Use 100 samples
chemtorch +experiment=graph data_module.subsample=100

Useful for debugging or testing configurations quickly.

Advanced Options#

parameter_limit

Limit the number of model parameters (for testing or architecture search). If the parameter limit is exceeded ChemTorch will skip the run.

  • Type: Integer or null

  • Default: null (no limit)

parameter_limit: 1000000  # Max 1M parameters

Common CLI Patterns#

Quick Testing#

# Fast test run: subsample data, run a single batch from every dataloader, no logging
chemtorch +experiment=graph \
    data_module.subsample=0.01 \
    +trainer.fast_dev_run=true \
    log=false

Inference#

# Load model and run predictions
chemtorch +experiment=graph \
    load_model=true \
    ckpt_path=path/to/checkpoint.ckpt \
    tasks=[predict] \
    predictions_save_path=predictions.csv

For advanced Hydra features, see Hydra Config Design or the official Hydra documentation.

Contents