Ask about the issue of units during the training process

[lingcon01]

Hello, I would like to ask about energy prediction using SCHNET. According to the results you provided, under the ST split of the tiny dataset, is its MAE 1.17 Hartree or 1.17 * 10^(-2) Hartree? I tried to reproduce the result, and I found that the MAE on the test set is around 0.75, but I am not sure about the unit. Could you clarify this?

[BerAnton]

Hello!
Could you please share run config file?

[lingcon01]

##config
`train:
batch_size: 128
seed: 2021
epochs: 500
num_workers: 0
restore_path: null
data_path: /home/suqun/data/nablaDFT/energy/tiny/train
data_name: train_2k_v2_formation_energy_w_forces
save_path: checkpoints/train/md17
log_interval: 10
lr: 0.0002
factor: 0.9
patience: 30
min_lr: 0.000001
energy_weight: 1
force_weight: 0
weight_decay: 1e-16

test:
test_interval: 1
test_batch_size: 100
data_path: /home/suqun/data/nablaDFT/energy/tiny/test
data_name: test_2k_conformers_v2_formation_energy_w_forces
`

##loss
loss_energy = loss_l1(pred, label) loss_force = loss_l1(dy, pdy)

##data_process
for db_row in tqdm(db.select(), total=len(db)): z = torch.from_numpy(db_row.numbers.copy()).long() positions = torch.from_numpy(db_row.positions.copy()).float() y = torch.from_numpy(np.array(db_row.data["energy"])).float() forces = torch.from_numpy(np.array(db_row.data["forces"])).float() molecule_size = len(positions) samples.append(Data(z=z, pos=positions, y=y, dy=forces, molecule_size=molecule_size))

##schnet
class SchNet(torch.nn.Module): r""" The re-implementation for SchNet from the "SchNet: A Continuous-filter Convolutional Neural Network for Modeling Quantum Interactions" https://arxiv.org/abs/1706.08566_ paper under the 3DGN gramework from "Spherical Message Passing for 3D Molecular Graphs" https://openreview.net/forum?id=givsRXsOt9r`_ paper.

    Args:
        energy_and_force (bool, optional): If set to :obj:`True`, will predict energy and take the negative of the derivative of the energy with respect to the atomic positions as predicted forces. (default: :obj:`False`)
        num_layers (int, optional): The number of layers. (default: :obj:`6`)
        hidden_channels (int, optional): Hidden embedding size. (default: :obj:`128`)
        out_channels (int, optional): Output embedding size. (default: :obj:`1`)
        num_filters (int, optional): The number of filters to use. (default: :obj:`128`)
        num_gaussians (int, optional): The number of gaussians :math:`\mu`. (default: :obj:`50`)
        cutoff (float, optional): Cutoff distance for interatomic interactions. (default: :obj:`10.0`).
"""

def __init__(self, energy_and_force=False, cutoff=10.0, num_layers=6, hidden_channels=128, out_channels=1,
             num_filters=128, num_gaussians=50):
    super(SchNet, self).__init__()

    self.energy_and_force = energy_and_force
    self.cutoff = cutoff
    self.num_layers = num_layers
    self.hidden_channels = hidden_channels
    self.out_channels = out_channels
    self.num_filters = num_filters
    self.num_gaussians = num_gaussians

    self.init_v = Embedding(100, hidden_channels)
    self.dist_emb = emb(0.0, cutoff, num_gaussians)

    self.update_vs = torch.nn.ModuleList([update_v(hidden_channels, num_filters) for _ in range(num_layers)])

    self.update_es = torch.nn.ModuleList([
        update_e(hidden_channels, num_filters, num_gaussians, cutoff) for _ in range(num_layers)])

    self.update_u = update_u(hidden_channels, out_channels)

    self.reset_parameters()

def reset_parameters(self):
    self.init_v.reset_parameters()
    for update_e in self.update_es:
        update_e.reset_parameters()
    for update_v in self.update_vs:
        update_v.reset_parameters()
    self.update_u.reset_parameters()

def forward(self, batch_data, mean, std):
    z, pos, batch = batch_data.z, batch_data.pos, batch_data.batch
    if self.energy_and_force:
        pos.requires_grad_()

    edge_index = radius_graph(pos, r=self.cutoff, batch=batch)
    row, col = edge_index
    dist = (pos[row] - pos[col]).norm(dim=-1)
    dist_emb = self.dist_emb(dist)

    v = self.init_v(z)

    for update_e, update_v in zip(self.update_es, self.update_vs):
        e = update_e(v, dist, dist_emb, edge_index)
        v = update_v(v, e, edge_index)
    u = self.update_u(v, batch)

    # calculate_forces:
    grad_outputs: List[Optional[torch.Tensor]] = [torch.ones_like(u)]
    forces = -torch.autograd.grad(
        outputs=[u],  # [n_graphs, ]
        inputs=[pos],  # [n_nodes, 3]
        grad_outputs=grad_outputs,
        retain_graph=True,  # Make sure the graph is not destroyed during training
        create_graph=True,  # Create graph for second derivative
        allow_unused=True,  # For complete dissociation turn to true
    )[
        0
    ]  # [n_nodes, 3]

    return u, forces`

[BerAnton]

First of all, model parameters differs from nablaDFT's SchNet.
Also you use conformers test split, not structures (ST).

Preferred way to reproduce results is to use run.py script with desired checkpoint and test data split.
Please refer to run configuration README for more details.

If you want to reproduce SchNet trained on tiny data split for structures (ST) test use this command:

python run.py --config schnet_test.yaml

schnet_test.yaml:

# Global variables
name: SchNet
dataset_name: dataset_test_structures
max_steps: 1000000
job_type: test
pretrained: SchNet_train_tiny # name of pretrained split or 'null'
ckpt_path: null # path to checkpoint for training resume or test run

# Datamodule parameters
root: ./datasets/nablaDFT/${.job_type}
batch_size: 32
num_workers: 8

# Devices
devices: [0]

# configs
defaults:
  - _self_
  - datamodule: nablaDFT_ase_test.yaml  # dataset config
  - model: schnet.yaml  # model config
  - callbacks: callbacks_spk.yaml  # pl callbacks config
  - loggers: wandb.yaml  # pl loggers config
  - trainer: test.yaml  # trainer config

# need this to set working dir as current dir
hydra:
  output_subdir: null
  run:
    dir: .
original_work_dir: ${hydra:runtime.cwd}

seed: 23

[lingcon01]

Thanks for your explanation and response, I understand. I have another question to ask. I used PygnablaDFT to read the dataset test_2k_conformers_v2_formation_energy_w_forces_test and saved it into a .pt file. Are the energy values in this dataset in Hartree units?

a = torch.load('test_2k_conformers_v2_formation_energy_w_forces_test.pt')
print(a[0].y)
tensor([-6.0566, -6.0559, -6.0675, ..., -7.9457, -7.4641, -7.4602])

[KuzmaKhrabrov]

All energy units in energy databases are Hartree units.

[lingcon01]

Thanks for your response! In the training process involving multiple partitioning methods for energy and force, is the loss calculation typically energy * 1 + force * 1 ?

[KuzmaKhrabrov]

In general, these are the hyperparameters of the training pipeline. For instance, we used 1:100 for training SchNet, PaiNN, GemNet. 1:1 for DimeNet and 2:100 for EquiformerV2

[lingcon01]

Thanks for your response! I would like to ask if this dataset contains a force label. "dataset_train_full": "https://a002dlils-kadurin-nabladft.obs.ru-moscow-1.hc.sbercloud.ru/data/nablaDFTv2/energy_databases/train_full_v2_formation_energy.db",

[KuzmaKhrabrov]

Thanks for your response! I would like to ask if this dataset contains a force label. "dataset_train_full": "https://a002dlils-kadurin-nabladft.obs.ru-moscow-1.hc.sbercloud.ru/data/nablaDFTv2/energy_databases/train_full_v2_formation_energy.db",

Please, use the updated dataset with force labels: https://a002dlils-kadurin-nabladft.obs.ru-moscow-1.hc.sbercloud.ru/data/nablaDFTv2/energy_databases/train_full_v2_formation_energy/train_full_v2_formation_energy_w_forces_wo_outliers.db