Skip to content

train_network.py: add debugging print for tensor and gradient norms #2136

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

Open
wants to merge 1 commit into
base: sd3
Choose a base branch
from
Open

train_network.py: add debugging print for tensor and gradient norms #2136

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
21 changes: 21 additions & 0 deletions train_network.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1430,6 +1430,24 @@ def restore_rng_state(rng_states: tuple[torch.ByteTensor, Optional[torch.ByteTen
if hasattr(network, "update_norms"):
network.update_norms()

if (args.debug_info_steps > 0) and (step % args.debug_info_steps == 0):
params = accelerator.unwrap_model(network).get_trainable_params()
grads, weights, numels = [], [], []
for p in params:
if p.requires_grad:
p_detached = p.detach()
weights.append(p_detached.norm(p=1).item())
if p.grad is not None:
grads.append(p.grad.detach().norm(p=1).item())
else:
grads.append(0.0)
numels.append(p_detached.numel())
total_grad = sum(grads) / sum(numels)
total_weight = sum(weights) / sum(numels)
accelerator.print(
f"\n[Step {step}] avr_grad={total_grad:.4E}, avr_weights={total_weight:.4E}"
)

optimizer.step()
lr_scheduler.step()
optimizer.zero_grad(set_to_none=True)
Expand Down Expand Up @@ -1722,6 +1740,9 @@ def setup_parser() -> argparse.ArgumentParser:
help="[EXPERIMENTAL] enable offloading of tensors to CPU during checkpointing for U-Net or DiT, if supported"
" / 勾配チェックポイント時にテンソルをCPUにオフロードする（U-NetまたはDiTのみ、サポートされている場合）",
)
parser.add_argument(
"--debug_info_steps", type=int, default=0, help="Log gradient/weight norms every N steps"
)
parser.add_argument(
"--no_metadata", action="store_true", help="do not save metadata in output model / メタデータを出力先モデルに保存しない"
)
Expand Down