feat: Add finetune method for MatterSim (#68)

* feat: Add finetune method for MatterSim

* Update pyproject.toml

pyproject.toml

@@ -49,6 +49,7 @@ dependencies = [
     "torchaudio>=2.2.0",
     "torchmetrics>=0.10.0",
     "torchvision>=0.17.0",
+    "wandb",
 ]
 
 [project.optional-dependencies]

script/finetune_mattersim.py

@@ -0,0 +1,248 @@
+# -*- coding: utf-8 -*-
+import argparse
+import os
+import pickle as pkl
+import random
+
+import numpy as np
+import torch
+import torch.distributed
+import wandb
+from ase.units import GPa
+
+from mattersim.datasets.utils.build import build_dataloader
+from mattersim.forcefield.m3gnet.scaling import AtomScaling
+from mattersim.forcefield.potential import Potential
+from mattersim.utils.atoms_utils import AtomsAdaptor
+from mattersim.utils.logger_utils import get_logger
+
+logger = get_logger()
+torch.distributed.init_process_group(backend="nccl")
+local_rank = int(os.environ["LOCAL_RANK"])
+
+
+def main(args):
+    args_dict = vars(args)
+    if args.wandb and local_rank == 0:
+        wandb_api_key = (
+            args.wandb_api_key
+            if args.wandb_api_key is not None
+            else os.getenv("WANDB_API_KEY")
+        )
+        wandb.login(key=wandb_api_key)
+        wandb.init(
+            project=args.wandb_project,
+            name=args.run_name,
+            config=args,
+            # id=args.run_name,
+            # resume="allow",
+        )
+
+    if args.wandb:
+        args_dict["wandb"] = wandb
+
+    torch.distributed.barrier()
+
+    # set random seed
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+
+    torch.cuda.set_device(local_rank)
+
+    if args.train_data_path.endswith(".pkl"):
+        with open(args.train_data_path, "rb") as f:
+            atoms_train = pkl.load(f)
+    else:
+        atoms_train = AtomsAdaptor.from_file(filename=args.train_data_path)
+    energies = []
+    forces = [] if args.include_forces else None
+    stresses = [] if args.include_stresses else None
+    logger.info("Processing training datasets...")
+    for atoms in atoms_train:
+        energies.append(atoms.get_potential_energy())
+        if args.include_forces:
+            forces.append(atoms.get_forces())
+        if args.include_stresses:
+            stresses.append(atoms.get_stress(voigt=False) / GPa)  # convert to GPa
+
+    dataloader = build_dataloader(
+        atoms_train,
+        energies,
+        forces,
+        stresses,
+        shuffle=True,
+        pin_memory=True,
+        is_distributed=True,
+        **args_dict,
+    )
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    # build energy normalization module
+    if args.re_normalize:
+        scale = AtomScaling(
+            atoms=atoms_train,
+            total_energy=energies,
+            forces=forces,
+            verbose=True,
+            **args_dict,
+        ).to(device)
+
+    if args.valid_data_path is not None:
+        if args.valid_data_path.endswith(".pkl"):
+            with open(args.valid_data_path, "rb") as f:
+                atoms_val = pkl.load(f)
+        else:
+            atoms_val = AtomsAdaptor.from_file(filename=args.train_data_path)
+        energies = []
+        forces = [] if args.include_forces else None
+        stresses = [] if args.include_stresses else None
+        logger.info("Processing validation datasets...")
+        for atoms in atoms_val:
+            energies.append(atoms.get_potential_energy())
+            if args.include_forces:
+                forces.append(atoms.get_forces())
+            if args.include_stresses:
+                stresses.append(atoms.get_stress(voigt=False) / GPa)  # convert to GPa
+        val_dataloader = build_dataloader(
+            atoms_val,
+            energies,
+            forces,
+            stresses,
+            pin_memory=True,
+            is_distributed=True,
+            **args_dict,
+        )
+    else:
+        val_dataloader = None
+
+    potential = Potential.from_checkpoint(
+        load_path=args.load_model_path,
+        load_training_state=False,
+        **args_dict,
+    )
+
+    if args.re_normalize:
+        potential.model.set_normalizer(scale)
+
+    potential.model = torch.nn.parallel.DistributedDataParallel(potential.model)
+    torch.distributed.barrier()
+
+    potential.train_model(
+        dataloader,
+        val_dataloader,
+        loss=torch.nn.HuberLoss(delta=0.01),
+        is_distributed=True,
+        **args_dict,
+    )
+
+    if local_rank == 0 and args.save_checkpoint:
+        wandb.save(os.path.join(args.save_path, "best_model.pth"))
+
+
+if __name__ == "__main__":
+    # Some important arguments
+    parser = argparse.ArgumentParser()
+
+    # path parameters
+    parser.add_argument(
+        "--run_name", type=str, default="example", help="name of the run"
+    )
+    parser.add_argument(
+        "--train_data_path", type=str, default="./sample.xyz", help="train data path"
+    )
+    parser.add_argument(
+        "--valid_data_path", type=str, default=None, help="valid data path"
+    )
+    parser.add_argument(
+        "--load_model_path",
+        type=str,
+        default="mattersim-v1.0.0-1m",
+        help="path to load the model",
+    )
+    parser.add_argument(
+        "--save_path", type=str, default="./results", help="path to save the model"
+    )
+    parser.add_argument(
+        "--save_checkpoint",
+        type=bool,
+        default=False,
+        action=argparse.BooleanOptionalAction,
+    )
+    parser.add_argument(
+        "--ckpt_interval",
+        type=int,
+        default=10,
+        help="save checkpoint every ckpt_interval epochs",
+    )
+    parser.add_argument("--device", type=str, default="cuda")
+
+    # model parameters
+    parser.add_argument("--cutoff", type=float, default=5.0, help="cutoff radius")
+    parser.add_argument(
+        "--threebody_cutoff",
+        type=float,
+        default=4.0,
+        help="cutoff radius for three-body term, which should be smaller than cutoff (two-body)",  # noqa: E501
+    )
+
+    # training parameters
+    parser.add_argument("--epochs", type=int, default=1000, help="number of epochs")
+    parser.add_argument("--batch_size", type=int, default=16)
+    parser.add_argument("--lr", type=float, default=2e-4)
+    parser.add_argument(
+        "--step_size",
+        type=int,
+        default=10,
+        help="step epoch for learning rate scheduler",
+    )
+    parser.add_argument(
+        "--include_forces",
+        type=bool,
+        default=True,
+        action=argparse.BooleanOptionalAction,
+    )
+    parser.add_argument(
+        "--include_stresses",
+        type=bool,
+        default=False,
+        action=argparse.BooleanOptionalAction,
+    )
+    parser.add_argument("--force_loss_ratio", type=float, default=1.0)
+    parser.add_argument("--stress_loss_ratio", type=float, default=0.1)
+    parser.add_argument("--early_stop_patience", type=int, default=10)
+    parser.add_argument("--seed", type=int, default=42)
+
+    # scaling parameters
+    parser.add_argument(
+        "--re_normalize",
+        type=bool,
+        default=False,
+        action=argparse.BooleanOptionalAction,
+        help="re-normalize the energy and forces according to the new data",
+    )
+    parser.add_argument("--scale_key", type=str, default="per_species_forces_rms")
+    parser.add_argument(
+        "--shift_key", type=str, default="per_species_energy_mean_linear_reg"
+    )
+    parser.add_argument("--init_scale", type=float, default=None)
+    parser.add_argument("--init_shift", type=float, default=None)
+    parser.add_argument(
+        "--trainable_scale",
+        type=bool,
+        default=False,
+        action=argparse.BooleanOptionalAction,
+    )
+    parser.add_argument(
+        "--trainable_shift",
+        type=bool,
+        default=False,
+        action=argparse.BooleanOptionalAction,
+    )
+
+    # wandb parameters
+    parser.add_argument("--wandb", action="store_true")
+    parser.add_argument("--wandb_api_key", type=str, default=None)
+    parser.add_argument("--wandb_project", type=str, default="wandb_test")
+    args = parser.parse_args()
+    main(args)

script/vasprun_to_xyz.py

@@ -0,0 +1,63 @@
+# -*- coding: utf-8 -*-
+import os
+import random
+
+from ase.io import write
+
+from mattersim.utils.atoms_utils import AtomsAdaptor
+
+vasp_files = [
+    "work/data/H/vasp/vasprun.xml",
+    "work/data/H/vasp_2/vasprun.xml",
+    "work/data/H/vasp_3/vasprun.xml",
+    "work/data/H/vasp_4/vasprun.xml",
+    "work/data/H/vasp_5/vasprun.xml",
+    "work/data/H/vasp_6/vasprun.xml",
+    "work/data/H/vasp_7/vasprun.xml",
+    "work/data/H/vasp_8/vasprun.xml",
+    "work/data/H/vasp_9/vasprun.xml",
+    "work/data/H/vasp_10/vasprun.xml",
+]
+train_ratio = 0.8
+validation_ratio = 0.1
+test_ratio = 0.1
+
+save_dir = "./xyz_files"
+os.makedirs(save_dir, exist_ok=True)
+
+
+def main():
+    atoms_train = []
+    atoms_validation = []
+    atoms_test = []
+
+    random.seed(42)
+
+    for vasp_file in vasp_files:
+        atoms_list = AtomsAdaptor.from_file(filename=vasp_file)
+        random.shuffle(atoms_list)
+        num_atoms = len(atoms_list)
+        num_train = int(num_atoms * train_ratio)
+        num_validation = int(num_atoms * validation_ratio)
+
+        atoms_train.extend(atoms_list[:num_train])
+        atoms_validation.extend(atoms_list[num_train : num_train + num_validation])
+        atoms_test.extend(atoms_list[num_train + num_validation :])
+
+    print(
+        f"Total number of atoms: {len(atoms_train) + len(atoms_validation) + len(atoms_test)}"  # noqa: E501
+    )
+
+    print(f"Number of atoms in the training set: {len(atoms_train)}")
+    print(f"Number of atoms in the validation set: {len(atoms_validation)}")
+    print(f"Number of atoms in the test set: {len(atoms_test)}")
+
+    # Save the training, validation, and test datasets to xyz files
+
+    write(f"{save_dir}/train.xyz", atoms_train)
+    write(f"{save_dir}/valid.xyz", atoms_validation)
+    write(f"{save_dir}/test.xyz", atoms_test)
+
+
+if __name__ == "__main__":
+    main()