Skip to content

fix small problem #2

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

fix small problem #2

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
40 changes: 22 additions & 18 deletions lib/vggflow/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -320,7 +320,7 @@ def _aggregate_and_log_metrics(self, global_step, num_processes):
aggregated_info[k] = torch.max(torch.stack(v))
elif '_all' in k:
aggregated_info[k] = torch.stack(v)
cache[k] = [torch.zeros_like(aggregated_info[k])] * num_processes
cache[k] = [torch.zeros_like(aggregated_info[k]) for _ in range(num_processes)]
else:
aggregated_info[k] = torch.mean(torch.stack(v))

Expand All @@ -334,45 +334,49 @@ def _aggregate_and_log_metrics(self, global_step, num_processes):
dist.all_reduce(v, op=dist.ReduceOp.MAX)
elif '_median' in k:
dist.all_gather(cache[k], v)
new_info[k.replace('_all', '')] = torch.median(cache[k][self.local_rank])
gathered = torch.cat([t.reshape(-1) for t in cache[k]], dim=0)
new_info[k.replace('_all', '')] = torch.median(gathered)
elif '_08quantile' in k:
dist.all_gather(cache[k], v)
new_info[k.replace('_all', '')] = torch.quantile(cache[k][self.local_rank], 0.8)
gathered = torch.cat([t.reshape(-1) for t in cache[k]], dim=0)
new_info[k.replace('_all', '')] = torch.quantile(gathered, 0.8)
elif '_std' in k:
dist.all_gather(cache[k], v)
new_info[k.replace('_all', '')] = torch.std(cache[k][self.local_rank])
gathered = torch.cat([t.reshape(-1) for t in cache[k]], dim=0)
new_info[k.replace('_all', '')] = torch.std(gathered)
else:
dist.all_reduce(v, op=dist.ReduceOp.SUM)

# Remove '_all' keys and add computed statistics
for k in list(aggregated_info.keys()):
# Average reduced means across processes; keep min/max as-is; drop raw "_all" tensors.
final_info = {}
for k, v in aggregated_info.items():
if '_all' in k:
aggregated_info.pop(k, None)
aggregated_info.update(new_info)
continue
if '_min' in k or '_max' in k:
final_info[k] = v
else:
final_info[k] = v / num_processes

# Average across processes
aggregated_info = {
k: v / num_processes if ('_min' not in k and '_max' not in k) else v
for k, v in aggregated_info.items()
}
# Add computed statistics (already global across all ranks; must not be divided)
final_info.update(new_info)

# Update rgrad threshold
rgrad_threshold = aggregated_info.get('rgrad_08quantile', 1.0).item()
rgrad_threshold = final_info.get('rgrad_08quantile', 1.0).item()

# Add epoch and step info
if self._is_main_process():
if self.scaler:
aggregated_info["grad_scale"] = self.scaler.get_scale()
aggregated_info["global_step"] = float(global_step)
final_info["grad_scale"] = self.scaler.get_scale()
final_info["global_step"] = float(global_step)

# Log to wandb
if self.config.logging.use_wandb:
import wandb
wandb.log(aggregated_info, step=global_step)
wandb.log(final_info, step=global_step)

# Log to console
self.logger.info(f"global_step={global_step} " +
" ".join([f"{k}={v:.6f}" for k, v in aggregated_info.items()]))
" ".join([f"{k}={v:.6f}" for k, v in final_info.items()]))

# Reset info dict
self.info = defaultdict(list)
Expand Down