Thanks to visit codestin.com
Credit goes to github.com

Skip to content

POC: enable to train at the double precision #207

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed

POC: enable to train at the double precision #207

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
223037f
Implement element filtering in the Ace datasets
Jul 19, 2023
32ce2dc
Make the Ace loader to load energy in float64
Jul 24, 2023
8ce19c6
Convert energy to right precision for training
Jul 24, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 10 additions & 4 deletions torchmdnet/datasets/ace.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,15 +15,17 @@ def __init__(
pre_transform=None,
pre_filter=None,
paths=None,
atomic_numbers=None,
max_gradient=None,
subsample_molecules=1,
):
assert isinstance(paths, (str, list))

arg_hash = f"{paths}{max_gradient}{subsample_molecules}"
arg_hash = f"{paths}{atomic_numbers}{max_gradient}{subsample_molecules}"
arg_hash = hashlib.md5(arg_hash.encode()).hexdigest()
self.name = f"{self.__class__.__name__}-{arg_hash}"
self.paths = paths
self.atomic_numbers = atomic_numbers
self.max_gradient = max_gradient
self.subsample_molecules = int(subsample_molecules)
super().__init__(root, transform, pre_transform, pre_filter)
Expand Down Expand Up @@ -180,6 +182,11 @@ def sample_iter(self, mol_ids=False):
fq = pt.tensor(mol["formal_charges"], dtype=pt.long)
q = fq.sum()

# Keep molecules with specific elements
if self.atomic_numbers:
if not set(z.numpy()).issubset(self.atomic_numbers):
continue
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This got mixed from #206, right?


for i_conf, (pos, y, neg_dy, pq, dp) in enumerate(load_confs(mol, n_atoms=len(z))):

# Skip samples with large forces
Expand Down Expand Up @@ -220,6 +227,7 @@ def processed_file_names(self):
def process(self):

print("Arguments")
print(f" atomic_numbers: {self.atomic_numbers}")
print(f" max_gradient: {self.max_gradient} eV/A")
print(f" subsample_molecules: {self.subsample_molecules}\n")

Expand Down Expand Up @@ -309,9 +317,7 @@ def get(self, idx):
atoms = slice(self.idx_mm[idx], self.idx_mm[idx + 1])
z = pt.tensor(self.z_mm[atoms], dtype=pt.long)
pos = pt.tensor(self.pos_mm[atoms], dtype=pt.float32)
y = pt.tensor(self.y_mm[idx], dtype=pt.float32).view(
1, 1
) # It would be better to use float64, but the trainer complaints
y = pt.tensor(self.y_mm[idx], dtype=pt.float64).view(1, 1)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would pass dtype as an argument to Ace here and store everything in the correct type. I do not see why store pos in float32 and y in float64.

neg_dy = pt.tensor(self.neg_dy_mm[atoms], dtype=pt.float32)
q = pt.tensor(self.q_mm[idx], dtype=pt.long)
pq = pt.tensor(self.pq_mm[atoms], dtype=pt.float32)
Expand Down
5 changes: 3 additions & 2 deletions torchmdnet/module.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,8 +127,9 @@ def step(self, batch, loss_fn, stage):
if batch.y.ndim == 1:
batch.y = batch.y.unsqueeze(1)

# y loss
loss_y = loss_fn(y, batch.y)
# y
y_dtype = {16: torch.float16, 32: torch.float32, 64: torch.float64}[self.hparams.precision]
loss_y = loss_fn(y, batch.y.to(y_dtype))
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How come you need this here but not a few lines above for neg_dy?


if stage in ["train", "val"] and self.hparams.ema_alpha_y < 1:
if self.ema[stage + "_y"] is None:
Expand Down
2 changes: 1 addition & 1 deletion torchmdnet/scripts/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@ def get_args():
parser.add_argument('--ema-alpha-neg-dy', type=float, default=1.0, help='The amount of influence of new losses on the exponential moving average of dy')
parser.add_argument('--ngpus', type=int, default=-1, help='Number of GPUs, -1 use all available. Use CUDA_VISIBLE_DEVICES=1, to decide gpus')
parser.add_argument('--num-nodes', type=int, default=1, help='Number of nodes')
parser.add_argument('--precision', type=int, default=32, choices=[16, 32], help='Floating point precision')
parser.add_argument('--precision', type=int, default=32, choices=[16, 32, 64], help='Floating point precision')
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

oh wow I totally missed this argument when I implemented #182

parser.add_argument('--log-dir', '-l', default='/tmp/logs', help='log file')
parser.add_argument('--splits', default=None, help='Npz with splits idx_train, idx_val, idx_test')
parser.add_argument('--train-size', type=number, default=None, help='Percentage/number of samples in training set (None to use all remaining samples)')
Expand Down