Skip to content
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

Update default hyperparams and rescaled losses for galaxies benchmark #16

Merged
merged 1 commit into from
Dec 5, 2024
Merged

Update default hyperparams and rescaled losses for galaxies benchmark #16

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
48 changes: 29 additions & 19 deletions benchmarks/galaxies/train_cosmology.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,8 +40,10 @@
from models.segnn import SEGNN
from models.nequip import NequIP
from models.diffpool import DiffPool
from models.pointnet import PointNet

from benchmarks.galaxies.dataset import get_halo_dataset, STD_PARAMS_DICT

from benchmarks.galaxies.dataset import get_halo_dataset

MLP_PARAMS = {
"feature_sizes": [128, 128, 128, 2],
Expand All @@ -50,7 +52,7 @@
GNN_PARAMS = {
"d_hidden": 128,
"n_layers": 3,
"message_passing_steps": 3,
"message_passing_steps": 8,
"message_passing_agg": "mean",
"activation": "gelu",
"norm": "none",
Expand All @@ -63,7 +65,7 @@
}

EGNN_PARAMS = {
"message_passing_steps": 3,
"message_passing_steps": 4,
"d_hidden": 128,
"n_layers": 3,
"activation": "gelu",
Expand All @@ -86,7 +88,7 @@
"combine_hierarchies_method": "mean",
"use_edge_features": True,
"task": "graph",
"mlp_readout_widths": [8, 2, 2],
"mlp_readout_widths": [8, 2],
}

POINTNET_PARAMS = {
Expand All @@ -97,19 +99,20 @@
"combine_hierarchies_method": "mean",
"use_edge_features": True,
"task": "graph",
"mlp_readout_widths": [8, 2, 2],
"mlp_readout_widths": [4, 2, 2],
"n_outputs": 2,
"k": 10,
}

SEGNN_PARAMS = {
"d_hidden": 128,
"n_layers": 3,
"message_passing_steps": 3,
"message_passing_steps": 8,
"message_passing_agg": "mean",
"scalar_activation": "gelu",
"gate_activation": "sigmoid",
"task": "graph",
"n_outputs": 2,
"output_irreps": e3nn.Irreps("1x0e"),
"readout_agg": "mean",
"mlp_readout_widths": (4, 2, 2),
"l_max_hidden": 2,
Expand All @@ -121,8 +124,7 @@
"d_hidden": 128,
"l_max":1,
"sphharm_norm": 'integral',
"irreps_out": e3nn.Irreps("1x0e"),
"message_passing_steps": 3,
"message_passing_steps": 4,
"n_layers": 3,
"message_passing_agg": "mean",
"readout_agg": "mean",
Expand Down Expand Up @@ -211,7 +213,6 @@ def __call__(self, x):


def loss_mse(pred_batch, cosmo_batch):
# return np.mean((pred_batch - cosmo_batch) ** 2)
return np.mean((pred_batch - cosmo_batch) ** 2, axis=0)


Expand Down Expand Up @@ -312,7 +313,7 @@ def run_expt(
use_tpcf="none",
n_steps=1000,
batch_size=32,
n_train=1248,
n_train=2048,
n_val=512,
n_test=512,
learning_rate=3e-4,
Expand Down Expand Up @@ -531,10 +532,14 @@ def run_expt(
# if early_stop.should_stop:
# print(f'Met early stopping criteria, breaking at epoch {step}')
# break

# Rescale losses by multiplying by variances, since we normalize during training
loss_rescaling = np.array([STD_PARAMS_DICT[param]**2 for param in target])

print(
"Training done.\n"
f"Final checkpoint test loss {test_loss_ckp}.\n"
f"Final rescaled test loss: {test_loss_ckp * loss_rescaling}.\n"
)

if plotting:
Expand Down Expand Up @@ -586,12 +591,16 @@ def main(model, feats, lr, decay, steps, batch_size, n_train, use_tpcf, k, data_
params = MLP_PARAMS
elif model == "GNN":
params = GNN_PARAMS
n_radial_basis = 64
elif model == "EGNN":
params = EGNN_PARAMS
n_radial_basis = 32
elif model == "SEGNN":
params = SEGNN_PARAMS
n_radial_basis = 32
elif model == "NequIP":
params = NEQUIP_PARAMS
n_radial_basis = 64
elif model == "DiffPool":
DIFFPOOL_PARAMS["gnn_kwargs"] = {
"d_hidden": 64,
Expand All @@ -602,14 +611,14 @@ def main(model, feats, lr, decay, steps, batch_size, n_train, use_tpcf, k, data_
DIFFPOOL_PARAMS["d_hidden"] = DIFFPOOL_PARAMS["gnn_kwargs"]["d_hidden"]
params = DIFFPOOL_PARAMS
elif model == "PointNet":
POINTNET_PARAMS["gnn_kwargs"] = {
"d_hidden": 64,
"n_layers": 4,
"message_passing_steps": 3,
"task": "node",
}
GNN_PARAMS["n_outputs"] = GNN_PARAMS["d_hidden"]
GNN_PARAMS["message_passing_steps"] = 3
GNN_PARAMS["n_layers"] = 4
GNN_PARAMS["task"] = "node"
POINTNET_PARAMS["gnn_kwargs"] = GNN_PARAMS
POINTNET_PARAMS["d_hidden"] = POINTNET_PARAMS["gnn_kwargs"]["d_hidden"]
params = POINTNET_PARAMS
n_radial_basis = 32
else:
raise NotImplementedError

Expand All @@ -623,7 +632,8 @@ def main(model, feats, lr, decay, steps, batch_size, n_train, use_tpcf, k, data_
n_steps=steps,
batch_size=batch_size,
n_train=n_train,
use_tpcf=use_tpcf
use_tpcf=use_tpcf,
n_radial_basis=n_radial_basis
)

if __name__ == "__main__":
Expand All @@ -636,7 +646,7 @@ def main(model, feats, lr, decay, steps, batch_size, n_train, use_tpcf, k, data_
parser.add_argument("--decay", type=float, help="Weight decay", default=1e-5)
parser.add_argument("--steps", type=int, help="Number of steps", default=5000)
parser.add_argument("--batch_size", type=int, help="Batch size", default=32)
parser.add_argument("--n_train", type=int, help="Number of training samples", default=1248)
parser.add_argument("--n_train", type=int, help="Number of training samples", default=2048)
parser.add_argument(
"--use_tpcf",
type=str,
Expand Down
2 changes: 1 addition & 1 deletion models/pointnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,7 @@ class PointNet(nn.Module):
def __call__(self, x, return_assignments=True):
# If graph prediction task, collect pooled embeddings at each hierarchy level
if self.task == "graph":
x_pool = jnp.zeros((self.n_downsamples, self.gnn_kwargs['d_output']))
x_pool = jnp.zeros((self.n_downsamples, self.gnn_kwargs['n_outputs']))

assignments = []
node_pos = x.nodes[..., :3]
Expand Down
Loading