smolvlm2-video-highlights2 / huggingface_exact_approach.py
avinashHuggingface108's picture
Fix deployment issues: permissions, short videos, and AI responses
bb657cb
#!/usr/bin/env python3
import os
import json
import tempfile
import torch
import warnings
from pathlib import Path
from transformers import AutoProcessor, AutoModelForImageTextToText
import subprocess
import logging
import argparse
from typing import List, Tuple, Dict
# Suppress warnings
os.environ["TOKENIZERS_PARALLELISM"] = "false"
warnings.filterwarnings("ignore", category=UserWarning)
warnings.filterwarnings("ignore", message=".*torchvision.*")
warnings.filterwarnings("ignore", message=".*torchcodec.*")
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
def get_video_duration_seconds(video_path: str) -> float:
"""Use ffprobe to get video duration in seconds."""
cmd = [
"ffprobe",
"-v", "quiet",
"-print_format", "json",
"-show_format",
video_path
]
result = subprocess.run(cmd, capture_output=True, text=True)
info = json.loads(result.stdout)
return float(info["format"]["duration"])
class VideoHighlightDetector:
def __init__(
self,
model_path: str,
device: str = None,
batch_size: int = 8
):
# Auto-detect device if not specified
if device is None:
if torch.cuda.is_available():
device = "cuda"
elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
self.device = device
self.batch_size = batch_size
# Initialize model and processor
self.processor = AutoProcessor.from_pretrained(model_path)
self.model = AutoModelForImageTextToText.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
# _attn_implementation="flash_attention_2"
).to(device)
# Store model path for reference
self.model_path = model_path
def analyze_video_content(self, video_path: str) -> str:
"""Analyze video content to determine its type and description."""
system_message = "You are a helpful assistant that can understand videos. Describe what type of video this is and what's happening in it."
messages = [
{
"role": "system",
"content": [{"type": "text", "text": system_message}]
},
{
"role": "user",
"content": [
{"type": "video", "path": video_path},
{"type": "text", "text": "What type of video is this and what's happening in it? Be specific about the content type and general activities you observe."}
]
}
]
inputs = self.processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt"
).to(self.device)
outputs = self.model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
return self.processor.decode(outputs[0], skip_special_tokens=True).lower().split("assistant: ")[1]
def determine_highlights(self, video_description: str, prompt_num: int = 1) -> str:
"""Determine what constitutes highlights based on video description with different prompts."""
system_prompts = {
1: "You are a highlight editor. List archetypal dramatic moments that would make compelling highlights if they appear in the video. Each moment should be specific enough to be recognizable but generic enough to potentially exist in other videos of this type.",
2: "You are a helpful visual-language assistant that can understand videos and edit. You are tasked helping the user to create highlight reels for videos. Highlights should be rare and important events in the video in question."
}
user_prompts = {
1: "List potential highlight moments to look for in this video:",
2: "List dramatic moments that would make compelling highlights if they appear in the video. Each moment should be specific enough to be recognizable but generic enough to potentially exist in any video of this type:"
}
messages = [
{
"role": "system",
"content": [{"type": "text", "text": system_prompts[prompt_num]}]
},
{
"role": "user",
"content": [{"type": "text", "text": f"""Here is a description of a video:\n\n{video_description}\n\n{user_prompts[prompt_num]}"""}]
}
]
print(f"Using prompt {prompt_num} for highlight detection")
inputs = self.processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt"
).to(self.device)
outputs = self.model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
response = self.processor.decode(outputs[0], skip_special_tokens=True)
# Extract the actual response with better formatting
if "Assistant: " in response:
clean_response = response.split("Assistant: ")[1]
elif "assistant: " in response.lower():
clean_response = response.lower().split("assistant: ")[1]
else:
# If no assistant tag found, try to extract meaningful content
parts = response.split("User:")
if len(parts) > 1:
clean_response = parts[-1].strip()
else:
clean_response = response
return clean_response.strip()
def process_segment(self, video_path: str, highlight_types: str) -> bool:
"""Process a video segment and determine if it contains highlights."""
messages = [
{
"role": "system",
"content": [{"type": "text", "text": "You are a STRICT video highlight analyzer. You must be very selective and only identify truly exceptional moments. Most segments should be rejected. Only select segments with high dramatic value, clear action, strong visual interest, or significant events. Be critical and selective."}]
},
{
"role": "user",
"content": [
{"type": "video", "path": video_path},
{"type": "text", "text": f"""Looking for these highlights:\n{highlight_types}\n\nDoes this video segment match ANY of these highlights?\n\nAnswer with ONE WORD ONLY:\nYES or NO\n\nNothing else. Just YES or NO."""}]
}
]
try:
inputs = self.processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt"
).to(self.device)
outputs = self.model.generate(
**inputs,
max_new_tokens=8, # Force very short responses
do_sample=False, # Use greedy decoding for consistency
temperature=0.1 # Very low temperature for strict adherence
)
response = self.processor.decode(outputs[0], skip_special_tokens=True)
# Extract assistant response
if "Assistant:" in response:
response = response.split("Assistant:")[-1].strip()
elif "assistant:" in response:
response = response.split("assistant:")[-1].strip()
response = response.lower()
print(f" πŸ€– AI Response: {response}")
# Simple yes/no detection - AI returns simple answers
response_clean = response.strip().replace("'", "").replace("-", "").replace(".", "").strip()
if response_clean.startswith("no"):
return False
elif response_clean.startswith("yes"):
return True
else:
# Default to no if unclear
return False
except Exception as e:
print(f" ❌ Error processing segment: {str(e)}")
return False
def _concatenate_scenes(
self,
video_path: str,
scene_times: list,
output_path: str,
with_effects: bool = True
):
"""Concatenate selected scenes into final video with optional effects."""
if not scene_times:
logger.warning("No scenes to concatenate, skipping.")
return
if with_effects:
self._concatenate_with_effects(video_path, scene_times, output_path)
else:
self._concatenate_basic(video_path, scene_times, output_path)
def _concatenate_basic(self, video_path: str, scene_times: list, output_path: str):
"""Basic concatenation without effects."""
filter_complex_parts = []
concat_inputs = []
for i, (start_sec, end_sec) in enumerate(scene_times):
filter_complex_parts.append(
f"[0:v]trim=start={start_sec}:end={end_sec},"
f"setpts=PTS-STARTPTS[v{i}];"
)
filter_complex_parts.append(
f"[0:a]atrim=start={start_sec}:end={end_sec},"
f"asetpts=PTS-STARTPTS[a{i}];"
)
concat_inputs.append(f"[v{i}][a{i}]")
concat_filter = f"{''.join(concat_inputs)}concat=n={len(scene_times)}:v=1:a=1[outv][outa]"
filter_complex = "".join(filter_complex_parts) + concat_filter
cmd = [
"ffmpeg",
"-y",
"-i", video_path,
"-filter_complex", filter_complex,
"-map", "[outv]",
"-map", "[outa]",
"-c:v", "libx264",
"-c:a", "aac",
output_path
]
logger.info(f"Running ffmpeg command: {' '.join(cmd)}")
subprocess.run(cmd, check=True, capture_output=True, text=True)
def _concatenate_with_effects(self, video_path: str, scene_times: list, output_path: str):
"""Concatenate with fade effects between segments."""
if len(scene_times) == 1:
# Single segment - just extract with fade in/out
start_sec, end_sec = scene_times[0]
duration = end_sec - start_sec
fade_duration = min(0.5, duration / 4) # 0.5s or 25% of duration, whichever is shorter
cmd = [
"ffmpeg", "-y",
"-i", video_path,
"-ss", str(start_sec),
"-t", str(duration),
"-vf", f"fade=in:0:{int(fade_duration*30)},fade=out:{int((duration-fade_duration)*30)}:{int(fade_duration*30)}",
"-af", f"afade=in:st=0:d={fade_duration},afade=out:st={duration-fade_duration}:d={fade_duration}",
"-c:v", "libx264", "-c:a", "aac",
output_path
]
else:
# Multiple segments - create with crossfade transitions
filter_parts = []
audio_parts = []
for i, (start_sec, end_sec) in enumerate(scene_times):
duration = end_sec - start_sec
fade_duration = min(0.3, duration / 6) # Shorter fades for multiple segments
# Video with fade
filter_parts.append(
f"[0:v]trim=start={start_sec}:end={end_sec},setpts=PTS-STARTPTS,"
f"fade=in:0:{int(fade_duration*30)},fade=out:{int((duration-fade_duration)*30)}:{int(fade_duration*30)}[v{i}]"
)
# Audio with fade
audio_parts.append(
f"[0:a]atrim=start={start_sec}:end={end_sec},asetpts=PTS-STARTPTS,"
f"afade=in:st=0:d={fade_duration},afade=out:st={duration-fade_duration}:d={fade_duration}[a{i}]"
)
# Concatenate all segments
video_concat = "".join([f"[v{i}]" for i in range(len(scene_times))])
audio_concat = "".join([f"[a{i}]" for i in range(len(scene_times))])
filter_complex = (
";".join(filter_parts) + ";" +
";".join(audio_parts) + ";" +
f"{video_concat}concat=n={len(scene_times)}:v=1:a=0[outv];" +
f"{audio_concat}concat=n={len(scene_times)}:v=0:a=1[outa]"
)
cmd = [
"ffmpeg", "-y",
"-i", video_path,
"-filter_complex", filter_complex,
"-map", "[outv]", "-map", "[outa]",
"-c:v", "libx264", "-c:a", "aac",
output_path
]
logger.info(f"Running ffmpeg command with effects: {' '.join(cmd)}")
result = subprocess.run(cmd, capture_output=True, text=True)
if result.returncode != 0:
logger.error(f"FFmpeg error: {result.stderr}")
# Fall back to basic concatenation
logger.info("Falling back to basic concatenation...")
self._concatenate_basic(video_path, scene_times, output_path)
def process_video(self, video_path: str, output_path: str, segment_length: float = 10.0, with_effects: bool = True) -> Dict:
"""Process video using exact HuggingFace approach."""
print("πŸš€ Starting HuggingFace Exact Video Highlight Detection")
print(f"πŸ“ Input: {video_path}")
print(f"πŸ“ Output: {output_path}")
print(f"⏱️ Segment Length: {segment_length}s")
print(f"🎨 With Effects: {with_effects}")
print()
# Get video duration
duration = get_video_duration_seconds(video_path)
if duration <= 0:
return {"error": "Could not determine video duration"}
print(f"πŸ“Ή Video duration: {duration:.1f}s ({duration/60:.1f} minutes)")
# Check if video is too short for meaningful highlights
if duration < segment_length * 2:
return {
"error": f"Video too short ({duration:.1f}s). Need at least {segment_length * 2:.1f}s for meaningful highlights.",
"video_description": "Video too short for analysis",
"total_segments": 0,
"selected_segments": 0
}
# Step 1: Analyze overall video content
print("🎬 Step 1: Analyzing overall video content...")
video_desc = self.analyze_video_content(video_path)
print(f"πŸ“ Video Description: {video_desc}")
print()
# Step 2: Get two different sets of highlights
print("🎯 Step 2: Determining highlight types (2 variations)...")
highlights1 = self.determine_highlights(video_desc, prompt_num=1)
highlights2 = self.determine_highlights(video_desc, prompt_num=2)
print(f"🎯 Highlight Set 1: {highlights1}")
print()
print(f"🎯 Highlight Set 2: {highlights2}")
print()
# Step 3: Split video into segments
temp_dir = os.path.join("/tmp", "temp_segments")
os.makedirs(temp_dir, mode=0o755, exist_ok=True)
kept_segments1 = []
kept_segments2 = []
segments_processed = 0
total_segments = int(duration / segment_length)
print(f"πŸ” Step 3: Processing {total_segments} segments of {segment_length}s each...")
for start_time in range(0, int(duration), int(segment_length)):
progress = int((segments_processed / total_segments) * 100) if total_segments > 0 else 0
end_time = min(start_time + segment_length, duration)
print(f"πŸ“Š Processing segment {segments_processed+1}/{total_segments} ({progress}%)")
print(f" ⏰ Time: {start_time}s - {end_time:.1f}s")
# Create segment
segment_path = f"{temp_dir}/segment_{start_time}.mp4"
cmd = [
"ffmpeg",
"-y",
"-v", "quiet", # Suppress FFmpeg output
"-i", video_path,
"-ss", str(start_time),
"-t", str(segment_length),
"-c:v", "libx264",
"-preset", "ultrafast", # Use ultrafast preset for speed
"-pix_fmt", "yuv420p", # Ensure compatible pixel format
segment_path
]
subprocess.run(cmd, check=True, capture_output=True)
# Process segment with both highlight sets
if self.process_segment(segment_path, highlights1):
print(" βœ… KEEPING SEGMENT FOR SET 1")
kept_segments1.append((start_time, end_time))
else:
print(" ❌ REJECTING SEGMENT FOR SET 1")
if self.process_segment(segment_path, highlights2):
print(" βœ… KEEPING SEGMENT FOR SET 2")
kept_segments2.append((start_time, end_time))
else:
print(" ❌ REJECTING SEGMENT FOR SET 2")
# Clean up segment file
os.remove(segment_path)
segments_processed += 1
print()
# Remove temp directory
os.rmdir(temp_dir)
# Calculate percentages of video kept for each highlight set
total_duration = duration
duration1 = sum(end - start for start, end in kept_segments1)
duration2 = sum(end - start for start, end in kept_segments2)
percent1 = (duration1 / total_duration) * 100
percent2 = (duration2 / total_duration) * 100
print(f"πŸ“Š Results Summary:")
print(f" 🎯 Highlight set 1: {percent1:.1f}% of video ({len(kept_segments1)} segments)")
print(f" 🎯 Highlight set 2: {percent2:.1f}% of video ({len(kept_segments2)} segments)")
# Choose the set with lower percentage unless it's zero
final_segments = kept_segments2 if (0 < percent2 <= percent1 or percent1 == 0) else kept_segments1
selected_set = "2" if final_segments == kept_segments2 else "1"
percent_used = percent2 if final_segments == kept_segments2 else percent1
print(f"πŸ† Selected Set {selected_set} with {len(final_segments)} segments ({percent_used:.1f}% of video)")
if not final_segments:
return {
"error": "No highlights detected in the video with either set of criteria",
"video_description": video_desc,
"highlights1": highlights1,
"highlights2": highlights2,
"total_segments": total_segments
}
# Step 4: Create final video
print(f"🎬 Step 4: Creating final highlights video...")
self._concatenate_scenes(video_path, final_segments, output_path, with_effects)
print("βœ… Highlights video created successfully!")
print(f"πŸŽ‰ SUCCESS! Created highlights with {len(final_segments)} segments")
print(f" πŸ“Ή Total highlight duration: {sum(end - start for start, end in final_segments):.1f}s")
print(f" πŸ“Š Percentage of original video: {percent_used:.1f}%")
# Return analysis results
return {
"success": True,
"video_description": video_desc,
"highlights1": highlights1,
"highlights2": highlights2,
"selected_set": selected_set,
"total_segments": total_segments,
"selected_segments": len(final_segments),
"selected_times": final_segments,
"total_duration": sum(end - start for start, end in final_segments),
"compression_ratio": percent_used / 100,
"output_path": output_path
}
def main():
parser = argparse.ArgumentParser(description='HuggingFace Exact Video Highlights')
parser.add_argument('video_path', help='Path to input video file')
parser.add_argument('--output', required=True, help='Path to output highlights video')
parser.add_argument('--save-analysis', action='store_true', help='Save analysis results to JSON')
parser.add_argument('--segment-length', type=float, default=10.0, help='Length of each segment in seconds (default: 10.0)')
parser.add_argument('--model', default='HuggingFaceTB/SmolVLM2-256M-Video-Instruct', help='SmolVLM2 model to use')
args = parser.parse_args()
# Validate input file
if not os.path.exists(args.video_path):
print(f"❌ Error: Video file not found: {args.video_path}")
return
# Create output directory if needed
output_dir = os.path.dirname(args.output)
if output_dir and not os.path.exists(output_dir):
os.makedirs(output_dir)
print(f"πŸš€ HuggingFace Exact SmolVLM2 Video Highlights")
print(f" Model: {args.model}")
print()
try:
# Initialize detector
print(f"πŸ”₯ Loading {args.model} for HuggingFace Exact Analysis...")
device = "mps" if torch.backends.mps.is_available() else ("cuda" if torch.cuda.is_available() else "cpu")
detector = VideoHighlightDetector(
model_path=args.model,
device=device,
batch_size=16
)
print("βœ… SmolVLM2 loaded successfully!")
print()
# Process video
results = detector.process_video(
video_path=args.video_path,
output_path=args.output,
segment_length=args.segment_length
)
# Save analysis if requested
if args.save_analysis:
analysis_file = args.output.replace('.mp4', '_exact_analysis.json')
with open(analysis_file, 'w') as f:
json.dump(results, f, indent=2, default=str)
print(f"πŸ“Š Analysis saved: {analysis_file}")
except Exception as e:
print(f"❌ Error: {str(e)}")
import traceback
traceback.print_exc()
if __name__ == "__main__":
main()