Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_schedulefree.py 3.67 KB
Newer Older
fanding2000's avatar
Format fix. More options in readme  
fanding2000 committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127 
import tempfile
from unittest.mock import MagicMock

import numpy as np
import pytest
import torch
import torch.nn.functional as F
from e3nn import o3

from mace import data, modules, tools
from mace.tools import scripts_utils, torch_geometric

try:
    import schedulefree
except ImportError:
    pytest.skip(
        "Skipping schedulefree tests due to ImportError", allow_module_level=True
    )

torch.set_default_dtype(torch.float64)

table = tools.AtomicNumberTable([6])
atomic_energies = np.array([1.0], dtype=float)
cutoff = 5.0


def create_mace(device: str, seed: int = 1702):
    torch_geometric.seed_everything(seed)

    model_config = {
        "r_max": cutoff,
        "num_bessel": 8,
        "num_polynomial_cutoff": 6,
        "max_ell": 3,
        "interaction_cls": modules.interaction_classes[
            "RealAgnosticResidualInteractionBlock"
        ],
        "interaction_cls_first": modules.interaction_classes[
            "RealAgnosticResidualInteractionBlock"
        ],
        "num_interactions": 2,
        "num_elements": 1,
        "hidden_irreps": o3.Irreps("8x0e + 8x1o"),
        "MLP_irreps": o3.Irreps("16x0e"),
        "gate": F.silu,
        "atomic_energies": atomic_energies,
        "avg_num_neighbors": 8,
        "atomic_numbers": table.zs,
        "correlation": 3,
        "radial_type": "bessel",
    }
    model = modules.MACE(**model_config)
    return model.to(device)


def create_batch(device: str):
    from ase import build

    size = 2
    atoms = build.bulk("C", "diamond", a=3.567, cubic=True)
    atoms_list = [atoms.repeat((size, size, size))]
    print("Number of atoms", len(atoms_list[0]))

    configs = [data.config_from_atoms(atoms) for atoms in atoms_list]
    data_loader = torch_geometric.dataloader.DataLoader(
        dataset=[
            data.AtomicData.from_config(config, z_table=table, cutoff=cutoff)
            for config in configs
        ],
        batch_size=1,
        shuffle=False,
        drop_last=False,
    )
    batch = next(iter(data_loader))
    batch = batch.to(device)
    batch = batch.to_dict()
    return batch


def do_optimization_step(
    model,
    optimizer,
    device,
):
    batch = create_batch(device)
    model.train()
    optimizer.train()
    optimizer.zero_grad()
    output = model(batch, training=True, compute_force=False)
    loss = output["energy"].mean()
    loss.backward()
    optimizer.step()
    model.eval()
    optimizer.eval()


@pytest.mark.parametrize("device", ["cpu", "cuda"])
def test_can_load_checkpoint(device):
    model = create_mace(device)
    optimizer = schedulefree.adamw_schedulefree.AdamWScheduleFree(model.parameters())
    args = MagicMock()
    args.optimizer = "schedulefree"
    args.scheduler = "ExponentialLR"
    args.lr_scheduler_gamma = 0.9
    lr_scheduler = scripts_utils.LRScheduler(optimizer, args)
    with tempfile.TemporaryDirectory() as d:
        checkpoint_handler = tools.CheckpointHandler(
            directory=d, keep=False, tag="schedulefree"
        )
        for _ in range(10):
            do_optimization_step(model, optimizer, device)
        batch = create_batch(device)
        output = model(batch)
        energy = output["energy"].detach().cpu().numpy()

        state = tools.CheckpointState(
            model=model, optimizer=optimizer, lr_scheduler=lr_scheduler
        )
        checkpoint_handler.save(state, epochs=0, keep_last=False)
        checkpoint_handler.load_latest(
            state=tools.CheckpointState(model, optimizer, lr_scheduler),
            swa=False,
        )
        batch = create_batch(device)
        output = model(batch)
        new_energy = output["energy"].detach().cpu().numpy()
        assert np.allclose(energy, new_energy, atol=1e-9)