checkpoints yay

This commit is contained in:
2025-07-26 17:11:07 -07:00
parent caa4ddb5e0
commit 262cb00b69
2 changed files with 283 additions and 18 deletions

View File

@@ -781,26 +781,55 @@ class VideoProcessor:
print(f"⚠️ Could not verify frame count: {e}")
def process_video(self) -> None:
"""Main video processing pipeline"""
"""Main video processing pipeline with checkpoint/resume support"""
self.processing_stats['start_time'] = time.time()
print("Starting VR180 video processing...")
# Load video info
self.load_video_info(self.config.input.video_path)
# Initialize checkpoint manager
from .checkpoint_manager import CheckpointManager
checkpoint_mgr = CheckpointManager(
self.config.input.video_path,
self.config.output.path
)
# Check for existing checkpoints
resume_info = checkpoint_mgr.get_resume_info()
if resume_info['can_resume']:
print(f"\n🔄 RESUME DETECTED:")
print(f" Found {resume_info['completed_chunks']} completed chunks")
print(f" Continue from where we left off? (saves time!)")
checkpoint_mgr.print_status()
# Calculate chunking parameters
chunk_size, overlap_frames = self.calculate_optimal_chunking()
# Calculate total chunks
total_chunks = 0
for _ in range(0, self.total_frames, chunk_size - overlap_frames):
total_chunks += 1
checkpoint_mgr.set_total_chunks(total_chunks)
# Process video in chunks
chunk_files = [] # Store file paths instead of frame data
temp_chunk_dir = Path(tempfile.mkdtemp(prefix="vr180_chunks_"))
try:
chunk_idx = 0
for start_frame in range(0, self.total_frames, chunk_size - overlap_frames):
end_frame = min(start_frame + chunk_size, self.total_frames)
frames_to_read = end_frame - start_frame
chunk_idx = len(chunk_files)
# Check if this chunk was already processed
existing_chunk = checkpoint_mgr.get_chunk_file(chunk_idx)
if existing_chunk:
print(f"\n✅ Chunk {chunk_idx} already processed: {existing_chunk.name}")
chunk_files.append(existing_chunk)
chunk_idx += 1
continue
print(f"\nProcessing chunk {chunk_idx}: frames {start_frame}-{end_frame}")
# Read chunk frames
@@ -818,7 +847,12 @@ class VideoProcessor:
chunk_path = temp_chunk_dir / f"chunk_{chunk_idx:04d}.npz"
print(f"Saving chunk {chunk_idx} to disk...")
np.savez_compressed(str(chunk_path), frames=matted_frames)
# Save to checkpoint
checkpoint_mgr.save_chunk(chunk_idx, None, source_chunk_path=chunk_path)
chunk_files.append(chunk_path)
chunk_idx += 1
# Free the frames from memory immediately
del matted_frames
@@ -837,21 +871,32 @@ class VideoProcessor:
if self.memory_manager.should_emergency_cleanup():
self.memory_manager.emergency_cleanup()
# Use streaming merge to avoid memory accumulation (fixes OOM)
print("\n🎬 Using streaming merge (no memory accumulation)...")
# Mark chunk processing as complete
checkpoint_mgr.mark_processing_complete()
# Determine audio source for final video
audio_source = None
if self.config.output.preserve_audio and Path(self.config.input.video_path).exists():
audio_source = self.config.input.video_path
# Stream merge chunks directly to output (no memory accumulation)
self.merge_chunks_streaming(
chunk_files=chunk_files,
output_path=self.config.output.path,
overlap_frames=overlap_frames,
audio_source=audio_source
)
# Check if merge was already done
if resume_info.get('merge_complete', False):
print("\n✅ Merge already completed in previous run!")
print(f" Output: {self.config.output.path}")
else:
# Use streaming merge to avoid memory accumulation (fixes OOM)
print("\n🎬 Using streaming merge (no memory accumulation)...")
# Determine audio source for final video
audio_source = None
if self.config.output.preserve_audio and Path(self.config.input.video_path).exists():
audio_source = self.config.input.video_path
# Stream merge chunks directly to output (no memory accumulation)
self.merge_chunks_streaming(
chunk_files=chunk_files,
output_path=self.config.output.path,
overlap_frames=overlap_frames,
audio_source=audio_source
)
# Mark merge as complete
checkpoint_mgr.mark_merge_complete()
print("✅ Streaming merge complete - no memory accumulation!")
@@ -869,11 +914,24 @@ class VideoProcessor:
print("Video processing completed!")
# Option to clean up checkpoints
print("\n🗄️ CHECKPOINT CLEANUP OPTIONS:")
print(" Checkpoints saved successfully and can be cleaned up")
print(" - Keep checkpoints for debugging: checkpoint_mgr.cleanup_checkpoints(keep_chunks=True)")
print(" - Remove all checkpoints: checkpoint_mgr.cleanup_checkpoints()")
print(f" - Checkpoint location: {checkpoint_mgr.checkpoint_dir}")
# For now, keep checkpoints by default (user can manually clean)
print("\n💡 Checkpoints kept for safety. Delete manually when no longer needed.")
finally:
# Clean up temporary chunk files
# Clean up temporary chunk files (but not checkpoints)
if temp_chunk_dir.exists():
print("Cleaning up temporary chunk files...")
shutil.rmtree(temp_chunk_dir)
try:
shutil.rmtree(temp_chunk_dir)
except Exception as e:
print(f"⚠️ Could not clean temp directory: {e}")
def _print_processing_statistics(self):
"""Print detailed processing statistics"""