smolvlm2-video-highlights / audio_enhanced_highlights_final.py
avinashHuggingface108's picture
Optimize: Visual-only mode for HuggingFace Spaces
5d1f54f
#!/usr/bin/env python3
"""
Audio-Enhanced Video Highlights Generator
Combines SmolVLM2 visual analysis with Whisper audio transcription
Supports 99+ languages including Telugu, Hindi, English
"""
import os
import sys
import cv2
import argparse
import json
import subprocess
import threading
import time
import tempfile
from pathlib import Path
from PIL import Image
from typing import List, Dict, Optional
import logging
# Add src directory to path for imports
sys.path.append(str(Path(__file__).parent / "src"))
try:
from src.smolvlm2_handler import SmolVLM2Handler
except ImportError:
print("❌ SmolVLM2Handler not found. Make sure to install dependencies first.")
sys.exit(1)
try:
import whisper
WHISPER_AVAILABLE = True
print("βœ… Whisper available for audio transcription")
except ImportError:
WHISPER_AVAILABLE = False
print("❌ Whisper not available. Install with: pip install openai-whisper")
sys.exit(1)
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class AudioVisualAnalyzer:
"""Comprehensive analyzer combining visual and audio analysis"""
def __init__(self, whisper_model_size="base", timeout_seconds=90, enable_visual=True, visual_only_mode=False):
"""Initialize with SmolVLM2 and optionally Whisper models"""
print("πŸ”§ Initializing Visual Analyzer...")
self.enable_visual = enable_visual
self.visual_only_mode = visual_only_mode
# Initialize SmolVLM2 for visual analysis
if self.enable_visual:
print("πŸ”₯ Loading SmolVLM2...")
self.vlm_handler = SmolVLM2Handler()
else:
print("πŸ”‡ Visual analysis disabled")
self.vlm_handler = None
self.timeout_seconds = timeout_seconds
# Skip Whisper loading in visual-only mode to save memory/resources
if self.visual_only_mode:
print("πŸ‘οΈ Visual-only mode enabled - skipping audio processing to optimize performance")
self.whisper_model = None
elif WHISPER_AVAILABLE:
print(f"πŸ“₯ Loading Whisper model ({whisper_model_size})...")
self.whisper_model = whisper.load_model(whisper_model_size)
print("βœ… Whisper model loaded successfully")
else:
self.whisper_model = None
print("⚠️ Whisper not available - audio analysis disabled")
def extract_audio_segments(self, video_path: str, segments: List[Dict]) -> List[str]:
"""Extract audio for specific video segments"""
audio_files = []
temp_dir = tempfile.mkdtemp()
for i, segment in enumerate(segments):
start_time = segment['start_time']
duration = segment['duration']
audio_path = os.path.join(temp_dir, f"segment_{i}.wav")
# Extract audio segment using FFmpeg
cmd = [
'ffmpeg', '-i', video_path,
'-ss', str(start_time),
'-t', str(duration),
'-vn', # No video
'-acodec', 'pcm_s16le', # Uncompressed audio
'-ar', '16000', # 16kHz sample rate for Whisper
'-ac', '1', # Mono
'-f', 'wav', # Force WAV format
'-y', # Overwrite
audio_path
]
try:
result = subprocess.run(cmd, check=True, capture_output=True, text=True)
if os.path.exists(audio_path) and os.path.getsize(audio_path) > 0:
audio_files.append(audio_path)
logger.info(f"πŸ“„ Extracted audio segment {i+1}: {duration:.1f}s")
else:
logger.warning(f"⚠️ Audio segment {i+1} is empty or missing")
audio_files.append(None)
except subprocess.CalledProcessError as e:
logger.warning(f"⚠️ No audio stream in segment {i+1} (this is normal for silent videos)")
audio_files.append(None)
return audio_files
def transcribe_audio_segment(self, audio_path: str) -> Dict:
"""Transcribe audio segment with Whisper"""
if not WHISPER_AVAILABLE or not audio_path or not os.path.exists(audio_path):
return {"text": "", "language": "unknown", "confidence": 0.0}
try:
result = self.whisper_model.transcribe(
audio_path,
language=None, # Auto-detect language
task="transcribe"
)
return {
"text": result.get("text", "").strip(),
"language": result.get("language", "unknown"),
"confidence": 1.0 # Whisper doesn't provide confidence scores
}
except Exception as e:
logger.error(f"❌ Audio transcription failed: {e}")
return {"text": "", "language": "unknown", "confidence": 0.0}
def analyze_visual_content(self, frame_path: str) -> Dict:
"""Analyze visual content using SmolVLM2 with robust error handling"""
# If visual analysis is disabled, return audio-focused fallback
if not self.enable_visual or self.vlm_handler is None:
logger.info("πŸ“Ή Visual analysis disabled, using audio-only mode")
return {"description": "Audio-only analysis mode - visual analysis disabled", "score": 7.0}
max_retries = 2
retry_count = 0
while retry_count < max_retries:
try:
def generate_with_timeout():
prompt = ("Analyze this video frame for interesting, engaging, or highlight-worthy content. "
"IMPORTANT: Start your response with 'Score: X/10' where X is a number from 1-10. "
"Then explain what makes it noteworthy. Focus on action, emotion, important moments, or visually striking elements. "
"Rate based on: Action/movement (high scores), People talking/interacting (medium-high), "
"Static scenes (low-medium), Boring/empty scenes (low scores).")
return self.vlm_handler.generate_response(frame_path, prompt)
# Run with timeout protection
thread_result = [None]
exception_result = [None]
def target():
try:
thread_result[0] = generate_with_timeout()
except Exception as e:
exception_result[0] = e
thread = threading.Thread(target=target)
thread.daemon = True
thread.start()
thread.join(self.timeout_seconds)
if thread.is_alive():
logger.warning(f"⏰ Visual analysis timed out after {self.timeout_seconds}s (attempt {retry_count + 1})")
retry_count += 1
if retry_count >= max_retries:
logger.info("πŸ”‡ Switching to audio-only mode due to visual timeout")
return {"description": "Visual analysis timed out - using audio-only mode", "score": 7.0}
continue
if exception_result[0]:
error_msg = str(exception_result[0])
if "probability tensor" in error_msg or "inf" in error_msg or "nan" in error_msg:
logger.warning(f"⚠️ Model inference error, retrying (attempt {retry_count + 1}): {error_msg}")
retry_count += 1
if retry_count >= max_retries:
return {"description": "Model inference failed after retries", "score": 6.0}
continue
else:
raise exception_result[0]
response = thread_result[0]
if not response or len(response.strip()) == 0:
logger.warning(f"⚠️ Empty response, retrying (attempt {retry_count + 1})")
retry_count += 1
if retry_count >= max_retries:
return {"description": "No meaningful response after retries", "score": 6.0}
continue
# Extract score from response
score = self.extract_score_from_text(response)
return {"description": response, "score": score}
except Exception as e:
error_msg = str(e)
logger.warning(f"⚠️ Visual analysis error (attempt {retry_count + 1}): {error_msg}")
retry_count += 1
if retry_count >= max_retries:
return {"description": f"Analysis failed after {max_retries} attempts: {error_msg}", "score": 6.0}
# Fallback if all retries failed
return {"description": "Analysis failed after all retry attempts", "score": 6.0}
def extract_score_from_text(self, text: str) -> float:
"""Extract numeric score from analysis text"""
import re
# Look for patterns like "Score: 8/10", "8/10", "score: 7", etc.
patterns = [
r'score:\s*(\d+(?:\.\d+)?)\s*/\s*10', # "Score: 8/10" (our new format)
r'(\d+(?:\.\d+)?)\s*/\s*10', # "8/10" or "7.5/10"
r'(?:score|rating|rate)(?:\s*[:=]\s*)(\d+(?:\.\d+)?)', # "score: 8" or "rating=7.5"
r'(\d+(?:\.\d+)?)\s*(?:out of|/)\s*10', # "8 out of 10"
r'(?:^|\s)(\d+(?:\.\d+)?)(?:\s*[/]\s*10)?(?:\s|$)', # Just numbers
]
for pattern in patterns:
matches = re.findall(pattern, text.lower())
if matches:
try:
score = float(matches[0])
return min(max(score, 1.0), 10.0) # Clamp between 1-10
except ValueError:
continue
return 6.0 # Default score if no pattern found
def calculate_combined_score(self, visual_score: float, audio_text: str, audio_lang: str) -> float:
"""Calculate combined score from visual and audio analysis"""
# Start with visual score
combined_score = visual_score
# Audio content scoring
if audio_text:
audio_bonus = 0.0
text_lower = audio_text.lower()
# Positive indicators
excitement_words = ['amazing', 'incredible', 'wow', 'fantastic', 'awesome', 'perfect', 'excellent']
action_words = ['goal', 'win', 'victory', 'success', 'breakthrough', 'achievement']
emotion_words = ['happy', 'excited', 'thrilled', 'surprised', 'shocked', 'love']
# Telugu positive indicators (basic)
telugu_positive = ['అద్భుఀం', 'చాలా బాగుంది', 'డాడ్', 'సూΰ°ͺర్']
# Count positive indicators
for word_list in [excitement_words, action_words, emotion_words, telugu_positive]:
for word in word_list:
if word in text_lower:
audio_bonus += 0.5
# Length bonus for substantial content
if len(audio_text) > 50:
audio_bonus += 0.3
elif len(audio_text) > 20:
audio_bonus += 0.1
# Language diversity bonus
if audio_lang in ['te', 'telugu']: # Telugu content
audio_bonus += 0.2
elif audio_lang in ['hi', 'hindi']: # Hindi content
audio_bonus += 0.2
combined_score += audio_bonus
# Clamp final score
return min(max(combined_score, 1.0), 10.0)
def analyze_segment(self, video_path: str, segment: Dict, temp_frame_path: str) -> Dict:
"""Analyze a single video segment with both visual and audio"""
start_time = segment['start_time']
duration = segment['duration']
logger.info(f"πŸ” Analyzing segment at {start_time:.1f}s ({duration:.1f}s duration)")
# Visual analysis
visual_analysis = self.analyze_visual_content(temp_frame_path)
# Skip audio analysis in visual-only mode to save resources
if self.visual_only_mode:
logger.info("πŸ‘οΈ Visual-only mode: skipping audio analysis")
audio_analysis = {"text": "", "language": "unknown", "confidence": 0.0}
# Use pure visual score for highlights
combined_score = visual_analysis['score']
else:
# Audio analysis
audio_files = self.extract_audio_segments(video_path, [segment])
audio_analysis = {"text": "", "language": "unknown", "confidence": 0.0}
if audio_files and audio_files[0]:
audio_analysis = self.transcribe_audio_segment(audio_files[0])
# Cleanup temporary audio file
try:
os.unlink(audio_files[0])
except:
pass
# Combined scoring
combined_score = self.calculate_combined_score(
visual_analysis['score'],
audio_analysis['text'],
audio_analysis['language']
)
return {
'start_time': start_time,
'duration': duration,
'visual_score': visual_analysis['score'],
'visual_description': visual_analysis['description'],
'audio_text': audio_analysis['text'],
'audio_language': audio_analysis['language'],
'combined_score': combined_score,
'selected': False
}
def extract_frames_at_intervals(video_path: str, interval_seconds: float = 10.0) -> List[Dict]:
"""Extract frames at regular intervals from video"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise ValueError(f"Cannot open video file: {video_path}")
fps = cap.get(cv2.CAP_PROP_FPS)
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
duration = total_frames / fps
logger.info(f"πŸ“Ή Video: {duration:.1f}s, {fps:.1f} FPS, {total_frames} frames")
segments = []
current_time = 0
while current_time < duration:
segment_duration = min(interval_seconds, duration - current_time)
segments.append({
'start_time': current_time,
'duration': segment_duration,
'frame_number': int(current_time * fps)
})
current_time += interval_seconds
cap.release()
return segments
def save_frame_at_time(video_path: str, time_seconds: float, output_path: str) -> bool:
"""Save a frame at specific time with robust frame extraction"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
return False
try:
fps = cap.get(cv2.CAP_PROP_FPS)
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
frame_number = int(time_seconds * fps)
# Ensure frame number is within valid range
frame_number = min(frame_number, total_frames - 1)
frame_number = max(frame_number, 0)
# Try to extract frame with fallback options
for attempt in range(3):
try:
# Try exact frame first
test_frame = frame_number + attempt
if test_frame >= total_frames:
test_frame = frame_number - attempt
if test_frame < 0:
test_frame = frame_number
cap.set(cv2.CAP_PROP_POS_FRAMES, test_frame)
ret, frame = cap.read()
if ret and frame is not None and frame.size > 0:
# Validate frame data
if len(frame.shape) == 3 and frame.shape[2] == 3: # Valid color frame
success = cv2.imwrite(output_path, frame)
if success:
cap.release()
return True
except Exception as e:
logger.warning(f"Frame extraction attempt {attempt + 1} failed: {e}")
continue
cap.release()
return False
except Exception as e:
logger.error(f"Critical error in frame extraction: {e}")
cap.release()
return False
def create_highlights_video(video_path: str, selected_segments: List[Dict], output_path: str):
"""Create highlights video from selected segments"""
if not selected_segments:
logger.error("❌ No segments selected for highlights")
return False
# Create temporary files for each segment
temp_files = []
temp_dir = tempfile.mkdtemp()
for i, segment in enumerate(selected_segments):
temp_file = os.path.join(temp_dir, f"segment_{i}.mp4")
cmd = [
'ffmpeg', '-i', video_path,
'-ss', str(segment['start_time']),
'-t', str(segment['duration']),
'-c', 'copy', # Copy streams without re-encoding
'-y', temp_file
]
try:
subprocess.run(cmd, check=True, capture_output=True)
temp_files.append(temp_file)
logger.info(f"βœ… Created segment {i+1}/{len(selected_segments)}")
except subprocess.CalledProcessError as e:
logger.error(f"❌ Failed to create segment {i+1}: {e}")
continue
if not temp_files:
logger.error("❌ No valid segments created")
return False
# Create concat file
concat_file = os.path.join(temp_dir, "concat.txt")
with open(concat_file, 'w') as f:
for temp_file in temp_files:
f.write(f"file '{temp_file}'\n")
# Concatenate segments
cmd = [
'ffmpeg', '-f', 'concat', '-safe', '0',
'-i', concat_file,
'-c', 'copy',
'-y', output_path
]
try:
subprocess.run(cmd, check=True, capture_output=True)
logger.info(f"βœ… Highlights video created: {output_path}")
# Cleanup
for temp_file in temp_files:
try:
os.unlink(temp_file)
except:
pass
try:
os.unlink(concat_file)
os.rmdir(temp_dir)
except:
pass
return True
except subprocess.CalledProcessError as e:
logger.error(f"❌ Failed to create highlights video: {e}")
return False
def main():
parser = argparse.ArgumentParser(description="Audio-Enhanced Video Highlights Generator")
parser.add_argument("video_path", help="Path to input video file")
parser.add_argument("--output", "-o", default="audio_enhanced_highlights.mp4",
help="Output highlights video path")
parser.add_argument("--interval", "-i", type=float, default=10.0,
help="Analysis interval in seconds (default: 10.0)")
parser.add_argument("--min-score", "-s", type=float, default=7.0,
help="Minimum score for highlights (default: 7.0)")
parser.add_argument("--max-highlights", "-m", type=int, default=5,
help="Maximum number of highlights (default: 5)")
parser.add_argument("--whisper-model", "-w", default="base",
choices=["tiny", "base", "small", "medium", "large"],
help="Whisper model size (default: base)")
parser.add_argument("--timeout", "-t", type=int, default=30,
help="Timeout for each analysis in seconds (default: 30)")
parser.add_argument("--save-analysis", action="store_true",
help="Save detailed analysis to JSON file")
args = parser.parse_args()
# Validate input
if not os.path.exists(args.video_path):
print(f"❌ Video file not found: {args.video_path}")
sys.exit(1)
print("🎬 Audio-Enhanced Video Highlights Generator")
print(f"πŸ“ Input: {args.video_path}")
print(f"πŸ“ Output: {args.output}")
print(f"⏱️ Analysis interval: {args.interval}s")
print(f"🎯 Minimum score: {args.min_score}")
print(f"πŸ† Max highlights: {args.max_highlights}")
print(f"πŸŽ™οΈ Whisper model: {args.whisper_model}")
print()
try:
# Initialize analyzer
analyzer = AudioVisualAnalyzer(
whisper_model_size=args.whisper_model,
timeout_seconds=args.timeout
)
# Extract segments for analysis
segments = extract_frames_at_intervals(args.video_path, args.interval)
print(f"πŸ“Š Analyzing {len(segments)} segments...")
analyzed_segments = []
temp_frame_path = "temp_frame.jpg"
for i, segment in enumerate(segments):
print(f"\nπŸ” Segment {i+1}/{len(segments)} (t={segment['start_time']:.1f}s)")
# Save frame for visual analysis
if save_frame_at_time(args.video_path, segment['start_time'], temp_frame_path):
# Analyze segment
analysis = analyzer.analyze_segment(args.video_path, segment, temp_frame_path)
analyzed_segments.append(analysis)
print(f" πŸ‘οΈ Visual: {analysis['visual_score']:.1f}/10")
print(f" πŸŽ™οΈ Audio: '{analysis['audio_text'][:50]}...' ({analysis['audio_language']})")
print(f" 🎯 Combined: {analysis['combined_score']:.1f}/10")
else:
print(f" ❌ Failed to extract frame")
# Cleanup temp frame
try:
os.unlink(temp_frame_path)
except:
pass
if not analyzed_segments:
print("❌ No segments analyzed successfully")
sys.exit(1)
# Select best segments
analyzed_segments.sort(key=lambda x: x['combined_score'], reverse=True)
selected_segments = [s for s in analyzed_segments if s['combined_score'] >= args.min_score]
selected_segments = selected_segments[:args.max_highlights]
print(f"\nπŸ† Selected {len(selected_segments)} highlights:")
for i, segment in enumerate(selected_segments):
print(f"{i+1}. t={segment['start_time']:.1f}s, score={segment['combined_score']:.1f}")
if segment['audio_text']:
print(f" Audio: \"{segment['audio_text'][:100]}...\"")
if not selected_segments:
print(f"❌ No segments met minimum score of {args.min_score}")
sys.exit(1)
# Create highlights video
print(f"\n🎬 Creating highlights video...")
success = create_highlights_video(args.video_path, selected_segments, args.output)
if success:
print(f"βœ… Audio-enhanced highlights created: {args.output}")
# Save analysis if requested
if args.save_analysis:
analysis_file = args.output.replace('.mp4', '_analysis.json')
with open(analysis_file, 'w') as f:
json.dump({
'input_video': args.video_path,
'output_video': args.output,
'settings': {
'interval': args.interval,
'min_score': args.min_score,
'max_highlights': args.max_highlights,
'whisper_model': args.whisper_model,
'timeout': args.timeout
},
'segments': analyzed_segments,
'selected_segments': selected_segments
}, f, indent=2)
print(f"πŸ“Š Analysis saved: {analysis_file}")
else:
print("❌ Failed to create highlights video")
sys.exit(1)
except KeyboardInterrupt:
print("\n⏹️ Operation cancelled by user")
sys.exit(1)
except Exception as e:
print(f"❌ Error: {e}")
sys.exit(1)
if __name__ == "__main__":
main()