Machine-Learning
/
aiuNN
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request 'updated memory fix' (#26) from feat/save_fix into main
Run VectorLoader Script / Explore-Gitea-Actions (push) Successful in 20s Details
Gitea Actions For AIIA / Explore-Gitea-Actions (push) Successful in 9m36s Details 

Reviewed-on: #26
This commit is contained in:
Falko Victor Habel 2025-07-03 11:50:07 +00:00
parent c92fa68e92 159ada872b
commit e7b9da37d6
1 changed files with 18 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
src/aiunn/finetune/memory_trainer.py
View File

@ -237,9 +237,10 @@ class MemoryOptimizedTrainer(aiuNNTrainer):
start_idx = start_batch if epoch == start_epoch else 0
progress_bar = tqdm(train_batches[start_idx:],
initial=start_idx,
total=len(train_batches),
desc=f"Epoch {epoch + 1}/{epochs}")
initial=start_idx,
total=len(train_batches),
desc=f"Epoch {epoch + 1}/{epochs}")
for batch_idx, (low_res, high_res) in progress_bar:
# Move data to device
@ -251,8 +252,10 @@ class MemoryOptimizedTrainer(aiuNNTrainer):
if hasattr(self, 'use_checkpointing') and self.use_checkpointing:
low_res.requires_grad_()
outputs = checkpoint(self.model, low_res)
outputs = outputs.clone() # <-- Clone added here
else:
outputs = self.model(low_res)
outputs = outputs.clone() # <-- Clone added here
loss = self.criterion(outputs, high_res)
# Scale loss for gradient accumulation
@ -266,8 +269,8 @@ class MemoryOptimizedTrainer(aiuNNTrainer):
# Update weights every accumulation_steps or at the end of epoch
should_step = (not self.use_gradient_accumulation or
(batch_idx + 1) % self.accumulation_steps == 0 or
batch_idx == len(train_batches) - 1)
(batch_idx + 1) % self.accumulation_steps == 0 or
batch_idx == len(train_batches) - 1)
if should_step:
self.scaler.step(self.optimizer)
@ -348,6 +351,16 @@ class MemoryOptimizedTrainer(aiuNNTrainer):
return self.best_loss
# Stop memory monitoring
if self.use_memory_profiling:
self.stop_monitoring = True
if self.memory_monitor_thread:
self.memory_monitor_thread.join(timeout=1)
print(f"Training completed. Peak GPU memory usage: {self.peak_memory:.2f}GB")
return self.best_loss
def get_memory_summary(self):
"""Get a summary of memory usage during training"""
if not self.memory_stats: