Source code for metatrain.cli.train

import argparse
import copy
import itertools
import logging
import os
import random
import re
import shutil
from pathlib import Path
from typing import Dict, List, Optional, Union

import numpy as np
import torch
from omegaconf import DictConfig, OmegaConf
from torch.utils.data import Subset

from metatrain.utils.data import Dataset

from ..utils.abc import ModelInterface, TrainerInterface
from ..utils.architectures import (
    check_architecture_options,
    get_default_hypers,
    import_architecture,
)
from ..utils.data import (
    DatasetInfo,
    TargetInfo,
    get_atomic_types,
    get_dataset,
    get_stats,
    load_indices,
)
from ..utils.data.dataset import _save_indices, _train_test_random_split
from ..utils.devices import pick_devices
from ..utils.distributed.logging import is_main_process
from ..utils.errors import ArchitectureError, OutOfMemoryError
from ..utils.io import (
    check_file_extension,
    load_model,
    model_from_checkpoint,
    trainer_from_checkpoint,
)
from ..utils.logging import WandbHandler, human_readable
from ..utils.omegaconf import (
    BASE_OPTIONS,
    check_units,
    expand_dataset_config,
    expand_loss_config,
)
from ..utils.pydantic import validate_base_options
from .eval import _eval_targets
from .export import _has_extensions
from .formatter import CustomHelpFormatter


def _is_indices_only_config(config: DictConfig) -> bool:
    """Check if config is indices-only (no systems/targets, just indices).

    This is used for validation_set and test_set configs that reference
    the training source file via indices only.

    :param config: The DictConfig to check.
    :return: True if config is indices-only, False otherwise.
    """
    return (
        isinstance(config, DictConfig)
        and "indices" in config
        and "systems" not in config
    )


def _validate_indices(indices: List[int], dataset_size: int, set_name: str) -> None:
    """Validate indices are within dataset bounds.

    :param indices: List of indices to validate
    :param dataset_size: Size of the dataset
    :param set_name: Name of the set (for error messages)
    :raises ValueError: If indices are out of range
    """
    if not indices:
        return  # Empty indices allowed (e.g., for empty val/test sets)
    max_idx = max(indices)
    if max_idx >= dataset_size:
        raise ValueError(
            f"{set_name} index {max_idx} out of range for "
            f"dataset of size {dataset_size}"
        )


def _add_train_model_parser(subparser: argparse._SubParsersAction) -> None:
    """Add `train_model` paramaters to an argparse (sub)-parser.

    :param subparser: The argparse (sub)-parser to add the parameters to.
    """

    if train_model.__doc__ is not None:
        description = train_model.__doc__.split(r":param")[0]
    else:
        description = None

    # If you change the synopsis of these commands or add new ones adjust the completion
    # script at `src/metatrain/share/metatrain-completion.bash`.
    parser = subparser.add_parser(
        "train",
        description=description,
        formatter_class=CustomHelpFormatter,
    )
    parser.set_defaults(callable="train_model")

    parser.add_argument(
        "options",
        type=str,
        help="Options YAML file for training.",
    )
    parser.add_argument(
        "-o",
        "--output",
        dest="output",
        type=str,
        required=False,
        default="model.pt",
        help="Path to save the final model (default: %(default)s).",
    )
    parser.add_argument(
        "-e",
        "--extensions",
        dest="extensions",
        type=str,
        required=False,
        default="extensions/",
        help=(
            "Folder where the extensions of the model, if any, will be collected "
            "(default: %(default)s)"
        ),
    )
    parser.add_argument(
        "--restart",
        dest="restart_from",
        type=_process_restart_from,
        required=False,
        help=(
            "Checkpoint file (.ckpt) to continue interrupted training. Set to `'auto'` "
            "to take the most recent checkpoint from the outputs directory."
        ),
    )
    parser.add_argument(
        "-r",
        "--override",
        dest="override_options",
        action="append",
        help="Command-line override flags.",
        default=[],
    )


def _prepare_train_model_args(args: argparse.Namespace) -> None:
    """Prepare arguments for train_model.

    :param args: The argparse.Namespace containing the arguments.
    """
    args.options = OmegaConf.load(args.options)
    # merge/override file options with command line options
    override_options = args.__dict__.pop("override_options")
    override_options = OmegaConf.from_dotlist(override_options)

    args.options = OmegaConf.merge(args.options, override_options)


def _setup_wandb_logging(logger: logging.Logger, args: argparse.Namespace) -> None:
    """Set up Weights and Biases logging if `wandb` options are provided.

    :param logger: The logger to use for logging messages.
    :param args: The argparse.Namespace containing the arguments.
    """
    # Don't modify the incoming args to prevent errors.
    detached_args = copy.deepcopy(args)

    # Be very graceful in parsing the options -- true errors should be handled once
    # the logger is setup in setup_logging!
    options = None
    try:
        _prepare_train_model_args(detached_args)
        options = detached_args.__dict__["options"]
    except Exception:
        ...

    # Try to setup wandb logging if the options are present.
    if options is not None and hasattr(options, "wandb") and is_main_process():
        try:
            import wandb
        except ImportError:
            raise ImportError(
                "Wandb is enabled but not installed. "
                "Please install wandb using `pip install wandb` to use this logger."
            )

        run = wandb.init(
            **options["wandb"], config=OmegaConf.to_container(options, resolve=True)
        )
        logger.addHandler(WandbHandler(run))


def _extend_wandb_config(options: Union[Dict, DictConfig]) -> None:
    """Update the wandb configuration with the expanded options.

    :param options: The options to update the configuration with.
    """
    try:
        import wandb

        if wandb.run is not None:
            wandb.config.update(
                OmegaConf.to_container(options, resolve=True), allow_val_change=True
            )
    except ImportError:
        pass


def _process_restart_from(restart_from: str) -> Optional[Union[str, Path]]:
    if restart_from != "auto":
        return restart_from

    pattern = re.compile(r".*\d{4}-\d{2}-\d{2}/\d{2}-\d{2}-\d{2}/*")
    checkpoints = sorted(
        # Use as_posix() so the forward-slash regex matches on Windows too
        (f for f in Path("outputs").glob("*/*/*.ckpt") if pattern.match(f.as_posix())),
        key=lambda f: f.stat().st_mtime,
        reverse=True,
    )

    return checkpoints[0] if checkpoints else None


[docs] def train_model( options: Union[DictConfig, Dict], output: Union[str, Path] = "model.pt", extensions: Union[str, Path] = "extensions/", checkpoint_dir: Union[str, Path] = ".", restart_from: Optional[Union[str, Path]] = None, ) -> None: """Train an atomistic machine learning model using provided ``options``. This function sets up the dataset and model architecture, then runs the training process. The dataset is prepared by reading structural data and target values from specified paths. The model architecture is dynamically imported and instantiated based on the configuration. Training is executed with the specified hyperparameters, and the trained model is saved to a designated output path. :param options: DictConfig containing the training options :param output: Path to save the final model :param extensions: Path to save the model extensions, if any :param checkpoint_dir: Path to save checkpoints and other intermediate output files like the fully expanded training options for a later restart. :param restart_from: File to continue training from. """ output = Path(check_file_extension(filename=output, extension=".pt")) extensions = Path(extensions) checkpoint_dir = Path(checkpoint_dir) ########################### # VALIDATE BASE OPTIONS ### ########################### # Training, test and validation set options are verified within the # `expand_dataset_config()` function. validate_base_options(OmegaConf.to_container(options)) ########################### # LOAD ARCHITECTURE ####### ########################### architecture_name = options["architecture"]["name"] try: architecture = import_architecture(architecture_name) except ImportError as e: raise ArchitectureError(e) from e logging.info(f"Running training for {architecture_name!r} architecture") Model = architecture.__model__ if not issubclass(Model, ModelInterface): raise TypeError( f"Model class for {architecture_name} must be a subclass of " " `metatrain.utils.abc.ModelInterface`" ) Trainer = architecture.__trainer__ if not issubclass(Trainer, TrainerInterface): raise TypeError( f"Trainer class for {architecture_name} must be a subclass of " " `metatrain.utils.abc.TrainerInterface`" ) ########################### # MERGE OPTIONS ########### ########################### options = OmegaConf.merge( BASE_OPTIONS, {"architecture": get_default_hypers(architecture_name)}, options, ) check_architecture_options( name=architecture_name, options=OmegaConf.to_container(options["architecture"]) ) ########################### # PROCESS BASE PARAMETERS # ########################### # process devices devices = pick_devices( architecture_devices=Model.__supported_devices__, desired_device=options["device"], ) # process base_precision/dtypes dtype = getattr(torch, f"float{options['base_precision']}") if dtype not in Model.__supported_dtypes__: raise ValueError( f"Requested dtype {dtype} is not supported. {architecture_name} only " f"supports {Model.__supported_dtypes__}." ) # process random seeds logging.info(f"Random seed of this run is {options['seed']}") torch.manual_seed(options["seed"]) np.random.seed(options["seed"]) random.seed(options["seed"]) os.environ["PYTHONHASHSEED"] = str(options["seed"]) if torch.cuda.is_available(): torch.cuda.manual_seed(options["seed"]) torch.cuda.manual_seed_all(options["seed"]) ############################ # SET UP TRAINING SET ###### ############################ logging.info("Setting up training set") options["training_set"] = expand_dataset_config(options["training_set"]) train_datasets = [] target_info_dict: Dict[str, TargetInfo] = {} extra_data_info_dict: Dict[str, TargetInfo] = {} for train_options in options["training_set"]: # loop over training sets dataset, target_info_dict_single, extra_data_info_dict_single = get_dataset( train_options ) train_datasets.append(dataset) intersecting_keys = target_info_dict.keys() & target_info_dict_single.keys() for key in intersecting_keys: if target_info_dict[key] != target_info_dict_single[key]: raise ValueError( f"Target information for key {key} differs between training sets. " f"Got {target_info_dict[key]} and {target_info_dict_single[key]}." ) target_info_dict.update(target_info_dict_single) intersecting_keys = ( extra_data_info_dict.keys() & extra_data_info_dict_single.keys() ) for key in intersecting_keys: if extra_data_info_dict[key] != extra_data_info_dict_single[key]: raise ValueError( f"Extra data information for key {key} differs between training " f"sets. Got {extra_data_info_dict[key]} and" f" {extra_data_info_dict_single[key]}." ) extra_data_info_dict.update(extra_data_info_dict_single) # Store full datasets before any index-based filtering # (needed when val/test use indices referencing training source) full_datasets = list(train_datasets) # Track explicit training indices (if specified in config) explicit_train_indices: List[Optional[List[int]]] = [None] * len(train_datasets) # Apply indices filtering to training set if specified for i_dataset, train_options in enumerate(options["training_set"]): if "indices" in train_options: train_idx = load_indices(train_options["indices"]) _validate_indices(train_idx, len(full_datasets[i_dataset]), "Training set") train_datasets[i_dataset] = Subset(full_datasets[i_dataset], train_idx) explicit_train_indices[i_dataset] = train_idx train_size = 1.0 ############################ # SET UP VALIDATION SET #### ############################ logging.info("Setting up validation set") val_datasets: List[Union[Dataset, Subset]] = [] train_indices: List[Optional[List[int]]] = [] val_indices: List[Optional[List[int]]] = [] val_is_indices_only = _is_indices_only_config(options["validation_set"]) if isinstance(options["validation_set"], (int, float)): val_size = options["validation_set"] train_size -= val_size for i_dataset, train_dataset in enumerate(train_datasets): train_dataset_new, val_dataset = _train_test_random_split( train_dataset=train_dataset, train_size=train_size, test_size=val_size, ) train_datasets[i_dataset] = train_dataset_new val_datasets.append(val_dataset) # Remap indices to original space if training_set had explicit indices orig_indices = explicit_train_indices[i_dataset] if orig_indices is not None: train_indices.append( [orig_indices[i] for i in train_dataset_new.indices] ) val_indices.append([orig_indices[i] for i in val_dataset.indices]) else: train_indices.append(train_dataset_new.indices) val_indices.append(val_dataset.indices) elif val_is_indices_only: # Indices-only config only works for single dataset if len(options["training_set"]) > 1: raise ValueError( "indices-only validation_set config is not supported with " "multi-dataset training_set. " "Use full config with systems/targets/indices for each dataset." ) # Indices reference the training source file(s) val_idx = load_indices(options["validation_set"]["indices"]) _validate_indices(val_idx, len(full_datasets[0]), "Validation set") # Create subset from full dataset (not the potentially filtered train_datasets) val_datasets.append(Subset(full_datasets[0], val_idx)) val_indices.append(val_idx) # Use explicit training indices if they were specified train_indices.append(explicit_train_indices[0]) else: options["validation_set"] = expand_dataset_config(options["validation_set"]) if len(options["validation_set"]) != len(options["training_set"]): raise ValueError( f"Validation dataset with length {len(options['validation_set'])} has " "a different size than the training datatset with length " f"{len(options['training_set'])}." ) check_units( actual_options=options["validation_set"], desired_options=options["training_set"], ) for valid_options in options["validation_set"]: dataset, _, _ = get_dataset(valid_options) # Apply indices filtering if specified in the config if "indices" in valid_options: idx = load_indices(valid_options["indices"]) _validate_indices(idx, len(dataset), "Validation set") dataset = Subset(dataset, idx) val_indices.append(idx) else: val_indices.append(None) val_datasets.append(dataset) train_indices.append(None) ############################ # SET UP TEST SET ########## ############################ logging.info("Setting up test set") test_datasets: List[Union[Dataset, Subset]] = [] test_indices: List[Optional[List[int]]] = [] test_is_indices_only = _is_indices_only_config(options.get("test_set")) if isinstance(options["test_set"], (int, float)): test_size = options["test_set"] train_size -= test_size for i_dataset, train_dataset in enumerate(train_datasets): train_dataset_new, test_dataset = _train_test_random_split( train_dataset=train_dataset, train_size=train_size, test_size=test_size, ) train_datasets[i_dataset] = train_dataset_new test_datasets.append(test_dataset) there_was_no_validation_split = train_indices[i_dataset] is None new_train_indices = ( train_dataset_new.indices if there_was_no_validation_split else [train_indices[i_dataset][i] for i in train_dataset_new.indices] # type: ignore[index] ) test_indices.append( test_dataset.indices if there_was_no_validation_split else [train_indices[i_dataset][i] for i in test_dataset.indices] # type: ignore[index] ) train_indices[i_dataset] = new_train_indices elif test_is_indices_only: # Indices-only config only works for single dataset if len(options["training_set"]) > 1: raise ValueError( "indices-only test_set config is not supported with " "multi-dataset training_set. " "Use full config with systems/targets/indices for each dataset." ) # Indices reference the training source file(s) test_idx = load_indices(options["test_set"]["indices"]) _validate_indices(test_idx, len(full_datasets[0]), "Test set") # Create subset from full dataset test_datasets.append(Subset(full_datasets[0], test_idx)) test_indices.append(test_idx) else: options["test_set"] = expand_dataset_config(options["test_set"]) if len(options["test_set"]) != len(options["training_set"]): raise ValueError( f"Test dataset with length {len(options['test_set'])} has a different " f"size than the training datatset with length " f"{len(options['training_set'])}." ) check_units( actual_options=options["test_set"], desired_options=options["training_set"], ) for test_options in options["test_set"]: dataset, _, _ = get_dataset(test_options) # Apply indices filtering if specified in the config if "indices" in test_options: idx = load_indices(test_options["indices"]) _validate_indices(idx, len(dataset), "Test set") dataset = Subset(dataset, idx) test_indices.append(idx) else: test_indices.append(None) test_datasets.append(dataset) # Expand loss options and finalize the hypers options = expand_loss_config(options) hypers = OmegaConf.to_container(options["architecture"], resolve=True) if is_main_process(): # Update the wandb configuration here because the initial wandb setup (in # `__main__.py`) only logs the unexpanded options. We pass the fully resolved # options (including BASE_OPTIONS) to the logger now. _extend_wandb_config(options) ############################################ # SAVE TRAIN, VALIDATION, TEST INDICES ##### ############################################ if is_main_process(): _save_indices(train_indices, val_indices, test_indices, checkpoint_dir) ########################### # CREATE DATASET_INFO ##### ########################### if options["architecture"].get("atomic_types") is None: # infer from datasets logging.info("Atomic types inferred from training and validation datasets") atomic_types = get_atomic_types(train_datasets + val_datasets) else: # use explicitly defined atomic types logging.info("Atomic types explicitly defined in options.yaml") atomic_types = sorted(options["architecture"]["atomic_types"]) logging.info(f"Model defined for atomic types: {atomic_types}") dataset_info = DatasetInfo( length_unit=options["training_set"][0]["systems"]["length_unit"], atomic_types=atomic_types, targets=target_info_dict, extra_data=extra_data_info_dict, ) ########################### # PRINT DATASET STATS ##### ########################### if sum(len(d) for d in train_datasets + val_datasets + test_datasets) < 1_000_000: # only print stats if the datasets are not too large (avoids hanging) _print_stats("Training", train_datasets, dataset_info) _print_stats("Validation", val_datasets, dataset_info) _print_stats("Test", test_datasets, dataset_info) else: logging.info( "Datasets are too large (>1M total structures) to calculate statistics " "quickly. Skipping statistics." ) ########################### # SAVE EXPANDED OPTIONS ### ########################### if is_main_process(): logging.info( "Restart options: " f"{checkpoint_dir.absolute().resolve() / 'options_restart.yaml'}" ) OmegaConf.save( config=options, f=checkpoint_dir / "options_restart.yaml", resolve=True, ) ########################### # SETTING UP MODEL ######## ########################### logging.info("Setting up model") # Resolving the model initialization options if restart_from is not None: training_context = "restart" elif "finetune" in hypers["training"]: if "read_from" not in hypers["training"]["finetune"]: raise ValueError( "Finetuning is enabled but no checkpoint was provided. Please " "provide one using the `read_from` option in the `finetune` " "section." ) restart_from = hypers["training"]["finetune"]["read_from"] training_context = "finetune" else: training_context = None try: if training_context == "restart" and restart_from is not None: logging.info(f"Restarting training from '{restart_from}'") checkpoint = torch.load( restart_from, weights_only=False, map_location="cpu" ) try: model = model_from_checkpoint(checkpoint, context="restart") except Exception as e: raise ValueError( f"The file {restart_from} does not contain a valid checkpoint for " f"the '{architecture_name}' architecture" ) from e model = model.restart(dataset_info) try: trainer = trainer_from_checkpoint( checkpoint=checkpoint, hypers=hypers["training"], context=training_context, # type: ignore ) except Exception as e: raise ValueError( f"The file {restart_from} does not contain a valid checkpoint for " f"the '{architecture_name}' trainer state" ) from e elif training_context == "finetune" and restart_from is not None: logging.info(f"Starting finetuning from '{restart_from}'") checkpoint = torch.load( restart_from, weights_only=False, map_location="cpu" ) try: model = model_from_checkpoint(checkpoint, context="finetune") except Exception as e: raise ValueError( f"The file {restart_from} does not contain a valid checkpoint for " f"the '{architecture_name}' architecture" ) from e model = model.restart(dataset_info) trainer = Trainer(hypers["training"]) else: logging.info("Starting training from scratch") model = Model(hypers["model"], dataset_info) trainer = Trainer(hypers["training"]) except Exception as e: raise ArchitectureError(e) from e n_params = sum(p.numel() for p in model.parameters() if p.requires_grad) logging.info( ( f"The model has {human_readable(n_params)} parameters " f"(actual number: {n_params})." ) ) ########################### # TRAIN MODEL ############# ########################### logging.info("Calling trainer") logging.info( "Intermediate checkpoints (if available): " f"{checkpoint_dir.absolute().resolve()}" ) try: trainer.train( model=model, dtype=dtype, devices=devices, train_datasets=train_datasets, val_datasets=val_datasets, checkpoint_dir=str(checkpoint_dir), ) except torch.cuda.OutOfMemoryError as e: raise ArchitectureError(OutOfMemoryError(e)) from e except Exception as e: raise ArchitectureError(e) from e if not is_main_process(): return # only save and evaluate on the main process ########################### # SAVE FINAL MODEL ######## ########################### logging.info("Training finished!") checkpoint_output = output.with_suffix(".ckpt") try: trainer.save_checkpoint(model, checkpoint_output) except Exception as e: raise ArchitectureError(e) if checkpoint_output.exists(): # Reload ensuring (best) model intended for inference model = load_model(checkpoint_output) logging.info(f"Final checkpoint: {checkpoint_output.absolute().resolve()}") mts_atomistic_model = model.export() # Final device could be different from devices[0] defined above in the case of # multi-GPU and/or distributed training final_device = next( itertools.chain( mts_atomistic_model.parameters(), mts_atomistic_model.buffers(), ) ).device # model is first saved and then reloaded 1) for good practice and 2) because # `AtomisticModel` only torchscripts (makes faster) during `save()` mts_atomistic_model.save( file=output, collect_extensions=extensions if _has_extensions() else None, ) logging.info(f"Exported model: {output.absolute().resolve()}") if extensions.exists(): logging.info(f"Extensions path: {extensions.absolute().resolve()}") if checkpoint_dir.absolute().resolve() != Path.cwd(): shutil.copy(output, checkpoint_dir / output) if checkpoint_output.exists(): shutil.copy(checkpoint_output, checkpoint_dir / checkpoint_output) ########################### # EVALUATE FINAL MODEL #### ########################### # TODO: possibly control this better if and when wrappers archs will be treated # differently in the future if architecture_name == "llpr": if not hypers["training"]["train_all_parameters"]: # Skip final evaluation for LLPR return mts_atomistic_model = load_model( path=output, extensions_directory=extensions if _has_extensions() else None, ) mts_atomistic_model = mts_atomistic_model.to(final_device) batch_size = _get_batch_size_from_hypers(hypers) if batch_size is None: logging.warning( "Could not find batch size in hypers dictionary. " "Using default value of 1 for final evaluation." ) batch_size = 1 else: logging.info(f"Running final evaluation with batch size {batch_size}") for i, train_dataset in enumerate(train_datasets): if len(train_datasets) == 1: extra_log_message = "" else: extra_log_message = f" with index {i}" logging.info(f"Evaluating training dataset{extra_log_message}") _eval_targets( mts_atomistic_model, train_dataset, dataset_info.targets, batch_size=batch_size, ) for i, val_dataset in enumerate(val_datasets): if len(val_datasets) == 1: extra_log_message = "" else: extra_log_message = f" with index {i}" logging.info(f"Evaluating validation dataset{extra_log_message}") _eval_targets( mts_atomistic_model, val_dataset, dataset_info.targets, batch_size=batch_size, ) for i, test_dataset in enumerate(test_datasets): if len(test_datasets) == 1: extra_log_message = "" else: extra_log_message = f" with index {i}" logging.info(f"Evaluating test dataset{extra_log_message}") _eval_targets( mts_atomistic_model, test_dataset, dataset_info.targets, batch_size=batch_size, )
def _get_batch_size_from_hypers(hypers: Union[Dict, DictConfig]) -> Optional[int]: # Recursively searches for "batch_size", "batch-size", "batch size", "batchsize", # or their upper-case equivalents in the hypers dictionary and returns the value. # If not found, it returns None. for key in hypers.keys(): value = hypers[key] if isinstance(value, dict) or isinstance(value, DictConfig): batch_size = _get_batch_size_from_hypers(value) if batch_size is not None: return batch_size if ( key.lower().replace("_", "").replace("-", "").replace(" ", "") == "batchsize" ): return value return None def _print_stats(name: str, datasets: List[Dataset], dataset_info: DatasetInfo) -> None: # Prints statistics about the datasets for i, dataset in enumerate(datasets): if len(datasets) == 1: index = "" else: index = f" {i}" logging.info(f"{name} dataset{index}:\n {get_stats(dataset, dataset_info)}")