app.py

from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
import uuid
from moviepy.editor import VideoFileClip
import os
import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import cv2
from ultralytics import YOLO
from heapq import heappush, heappushpop
import numpy as np
import uuid
import uuid
from ultralytics import YOLO
import gradio as gr

# # default: Load the model on the available device(s)
# model = Qwen2VLForConditionalGeneration.from_pretrained(
#     "Qwen/Qwen2-VL-7B-Instruct", torch_dtype="auto", device_map="auto"
# )

# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-7B-Instruct",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processer
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")

# The default range for the number of visual tokens per image in the model is 4-16384. You can set min_pixels and max_pixels according to your needs, such as a token count range of 256-1280, to balance speed and memory usage.
# min_pixels = 256*28*28
# max_pixels = 1280*28*28
# processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels)

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "openai/whisper-large-v3"

model_whisper = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model_whisper.to(device)

processor_whisper = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model_whisper,
    tokenizer=processor_whisper.tokenizer,
    feature_extractor=processor_whisper.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
    return_timestamps=True
)

output_directory = "temp"  # Replace with your desired output directory
os.makedirs(output_directory, exist_ok=True)

def extract_audio(video_path):
    try:
        # Load the video file
        video = VideoFileClip(video_path)

        # Extract the audio
        audio = video.audio

        # Generate a unique filename using uuid
        unique_filename = f"{uuid.uuid4()}.mp3"
        audio_output_path = f"{output_directory}/{unique_filename}"

        # Save the audio to the unique file
        audio.write_audiofile(audio_output_path)

        result = pipe(audio_output_path)

        os.remove(audio_output_path)

        return result["text"]

    except Exception as e:

        print(f"Error: {str(e)}")

        return ""

output_dir = '/content/images'
model_yolo = YOLO('/model/best.pt')

def extract_top_weapon_frames(video_path, threshold=30):
    os.makedirs(output_dir, exist_ok=True)
    saved_paths = {
        'original': [],  # Paths for original frames
        'boxed': []     # Paths for frames with boxes
    }


    weapon_classes = ['weapon', 'knife']
    top_frames = []  # (confidence_score, original_frame, boxed_frame, frame_number)

    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        print("Error: Could not open video.")
        return saved_paths

    ret, prev_frame = cap.read()
    if not ret:
        print("Error: Could not read the first frame.")
        return saved_paths

    prev_gray = cv2.cvtColor(prev_frame, cv2.COLOR_BGR2GRAY)
    frame_number = 0

    while True:
        ret, frame = cap.read()
        if not ret:
            break

        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
        frame_diff = cv2.absdiff(gray, prev_gray)
        mean_diff = frame_diff.mean()

        if mean_diff > threshold:
            print(f"Processing frame {frame_number}")
            results = model_yolo.predict(source=frame, show=False)

            frame_max_conf = 0
            frame_with_boxes = frame.copy()

            for result in results:
                for box in result.boxes:
                    class_id = int(box.cls[0])
                    class_name = model_yolo.names[class_id]
                    confidence = float(box.conf[0])

                    if class_name in weapon_classes:
                        frame_max_conf = max(frame_max_conf, confidence)
                        x1, y1, x2, y2 = map(int, box.xyxy[0])
                        cv2.rectangle(frame_with_boxes, (x1, y1), (x2, y2), (0, 255, 0), 2)
                        label = f"{class_name} ({confidence:.2f})"
                        cv2.putText(frame_with_boxes, label, (x1, y1 - 10),
                                  cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

            if frame_max_conf > 0:
                if len(top_frames) < 2:
                    heappush(top_frames, (frame_max_conf, frame.copy(), frame_with_boxes, frame_number))
                elif frame_max_conf > top_frames[0][0]:
                    heappushpop(top_frames, (frame_max_conf, frame.copy(), frame_with_boxes, frame_number))

        prev_gray = gray
        frame_number += 1

    # Save the top 2 frames (both original and with boxes)
    for confidence, original_frame, boxed_frame, _ in sorted(top_frames, reverse=True):
        # Save original frame
        original_filename = f"{uuid.uuid4()}.jpg"
        original_path = os.path.join(output_dir, original_filename)
        cv2.imwrite(original_path, original_frame)
        saved_paths['original'].append(original_path)

        # Save frame with boxes
        boxed_filename = f"{uuid.uuid4()}.jpg"
        boxed_path = os.path.join(output_dir, boxed_filename)
        cv2.imwrite(boxed_path, boxed_frame)
        saved_paths['boxed'].append(boxed_path)

        print(f"Saved frame pair with confidence {confidence:.3f}")

    cap.release()
    return saved_paths

def detect_weapon_image(source_image_path):

    # Ensure the output directory exists
    os.makedirs(output_dir, exist_ok=True)

    # Run YOLO predictions
    results = model_yolo.predict(source=source_image_path, save=False, show=False)

    # List to store paths to saved images
    saved_paths = []

    for result in results:
        # Get the annotated image
        annotated_img = result.plot()

        # Generate a unique filename using UUID
        unique_filename = f"{uuid.uuid4()}.jpg"
        output_path = os.path.join(output_dir, unique_filename)

        # Save the annotated image
        cv2.imwrite(output_path, annotated_img)
        saved_paths.append(output_path)

    return saved_paths
def response(messages):
  # Preparation for inference
  text = processor.apply_chat_template(
      messages, tokenize=False, add_generation_prompt=True
  )
  image_inputs, video_inputs = process_vision_info(messages)
  inputs = processor(
      text=[text],
      images=image_inputs,
      videos=video_inputs,
      padding=True,
      return_tensors="pt",
  )
  inputs = inputs.to("cuda")

  # Inference: Generation of the output
  generated_ids = model.generate(**inputs, max_new_tokens=1024)
  generated_ids_trimmed = [
      out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
  ]
  output_text = processor.batch_decode(
      generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
  )
  return output_text[0]


system_prompt = """
Analyze the image for illegal items or contraband. Detect and categorize objects like guns, knives, drugs, and hidden compartments. Highlight areas of interest and provide:

1. A detailed explanation in Thai describing illegal items and their context.
2. A JSON output summarizing the findings.

Output Example:
1. Explanation (Thai): (detailed explanation in Thai describing illegal items and their context.)
2. JSON: [{"category": "weapon", "type": "gun"}]
"""


def is_mp4_file(file_path):
    return os.path.isfile(file_path) and file_path.lower().endswith(".mp4")

def process_inputs(text_input, file_input):

    if is_mp4_file(file_input):
        extract_images_from_video = extract_top_weapon_frames(file_input)
        transcription = extract_audio(file_input)

        try:
            # Prepare image content for messages
            image_content = []

            # Check if we have any original images
            if extract_images_from_video['original']:
                # Add first image if available
                image_content.append({
                    "type": "image",
                    "image": f"file://{extract_images_from_video['original'][0]}"
                })

                # Add second image if available
                if len(extract_images_from_video['original']) > 1:
                    image_content.append({
                        "type": "image",
                        "image": f"file://{extract_images_from_video['original'][1]}"
                    })

            # Create messages list with available content
            messages = [{"role": "system", "content": system_prompt},
                {
                    "role": "user",
                    "content": [
                        *image_content,  # Unpack available image content
                        {"type": "text", "text": f"Content From Social Media Post: {text_input}."},
                        {"type": "text", "text": f"this is transcription from video:{transcription}"}
                    ]
                }
            ]

            # Return response and available boxed images (empty list if none)
            result = response(messages), extract_images_from_video.get('boxed', [])
            return result

        except Exception as e:
            return f"Error: {str(e)}", []


    else:
      try:
        # Call your response function with text and file path
        messages = [ {"role": "system", "content": system_prompt},

        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": f"file://{file_input}",
                },
                {"type": "text", "text": f"Content From Social Media Post: {text_input}."},
            ],
        }]

        result = response(messages)
        detect_weapon = detect_weapon_image(file_input)
        # Optionally, delete the temporary file after processing

        return result,detect_weapon
      except Exception as e:
          # Handle any exceptions and return the error
        return f"Error: {str(e)}",[]

# Create the Gradio interface
demo = gr.Interface(
    fn=process_inputs,
    inputs=[
        gr.Textbox(
            label="Text Input",
            placeholder="Enter your text here...",
            lines=3
        ),
        gr.File(
            label="File Upload",
            file_types=[".mp4", ".png", ".jpeg",".jpg"],
            type="filepath"
        )
    ],
    outputs= [gr.Textbox(label="Process Results", lines=8),
              gr.Gallery(label="Generated images", show_label=False, elem_id="gallery", columns=[2], rows=[1], object_fit="contain", height="auto")],

    title="Text and File Input Processor Qwen2-VL-7B-Instruct",
    description="Enter text and/or upload a file to process them together",
)

if __name__ == "__main__":
    demo.launch()