Update app.py

app.py

@@ -3,7 +3,7 @@ import random
 import uuid
 import json
 import time
-import asyncio
+import re
 from threading import Thread
 
 import gradio as gr
@@ -12,7 +12,7 @@ import torch
 import numpy as np
 from PIL import Image
 import cv2
-import requests
+import supervision as sv
 
 from transformers import (
     Qwen3VLMoeForConditionalGeneration,
@@ -50,7 +50,7 @@ processor_q3vl = AutoProcessor.from_pretrained(MODEL_ID_Q3VL, trust_remote_code=
 model_q3vl = Qwen3VLMoeForConditionalGeneration.from_pretrained(
     MODEL_ID_Q3VL,
     trust_remote_code=True,
-    dtype=torch.float16
+    torch_dtype=torch.float16  # Corrected from 'dtype' to 'torch_dtype'
 ).to(device).eval()
 
 
@@ -93,7 +93,6 @@ def generate_image(text: str, image: Image.Image,
     messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": text}]}]
     prompt_full = processor_q3vl.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     
-    # FIX: Removed truncation=True and max_length to prevent the ValueError
     inputs = processor_q3vl(
         text=[prompt_full], images=[image], return_tensors="pt", padding=True
     ).to(device)
@@ -129,14 +128,12 @@ def generate_video(text: str, video_path: str,
 
     messages = [{"role": "user", "content": [{"type": "text", "text": text}]}]
     images_for_processor = []
-    # Add an <|image|> placeholder for each frame in the message
     for frame, timestamp in frames_with_ts:
-        messages[0]["content"].insert(0, {"type": "image"}) # Insert at beginning to match common patterns
+        messages[0]["content"].insert(0, {"type": "image"})
         images_for_processor.append(frame)
 
     prompt_full = processor_q3vl.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
 
-    # FIX: Removed truncation=True and max_length to prevent the ValueError
     inputs = processor_q3vl(
         text=[prompt_full], images=images_for_processor, return_tensors="pt", padding=True
     ).to(device)
@@ -156,71 +153,198 @@ def generate_video(text: str, video_path: str,
         time.sleep(0.01)
         yield buffer, buffer
 
+# --- Object Detection & Pointing Functions ---
+
+def parse_model_output_for_coords(text_output, task_type):
+    """Parses text to find normalized coordinates using regex and json."""
+    match = re.search(r'\[\[.*?\]\]', text_output)
+    if not match:
+        return []
+
+    try:
+        coords_str = match.group(0)
+        coords = json.loads(coords_str)
+        
+        if not isinstance(coords, list): return []
+        if task_type == "Object Detection":
+            return [c for c in coords if isinstance(c, list) and len(c) == 4 and all(isinstance(n, (int, float)) for n in c)]
+        elif task_type == "Point Detection":
+            return [c for c in coords if isinstance(c, list) and len(c) == 2 and all(isinstance(n, (int, float)) for n in c)]
+        return []
+    except (json.JSONDecodeError, TypeError):
+        return []
+
+def create_annotated_image_from_normalized(image, bboxes_normalized, object_name="Object"):
+    """Draws bounding boxes on an image from normalized coordinates."""
+    if not bboxes_normalized: return image
+    
+    original_width, original_height = image.size
+    annotated_image = np.array(image.convert("RGB"))
+    
+    bboxes_absolute = []
+    for bbox in bboxes_normalized:
+        x_min, y_min, x_max, y_max = bbox
+        bboxes_absolute.append([
+            int(x_min * original_width), int(y_min * original_height),
+            int(x_max * original_width), int(y_max * original_height)
+        ])
+    
+    detections = sv.Detections(xyxy=np.array(bboxes_absolute, dtype=np.float32))
+    bounding_box_annotator = sv.BoxAnnotator(thickness=2)
+    label_annotator = sv.LabelAnnotator(text_thickness=1, text_scale=0.5)
+    labels = [f"{object_name} #{i+1}" for i in range(len(detections))]
+    
+    annotated_image = bounding_box_annotator.annotate(scene=annotated_image, detections=detections)
+    annotated_image = label_annotator.annotate(scene=annotated_image, detections=detections, labels=labels)
+    
+    return Image.fromarray(annotated_image)
+
+def create_point_annotated_image_from_normalized(image, points_normalized):
+    """Draws points on an image from normalized coordinates."""
+    if not points_normalized: return image
+    
+    original_width, original_height = image.size
+    annotated_image = np.array(image.convert("RGB"))
+    
+    points_absolute = [[int(p[0] * original_width), int(p[1] * original_height)] for p in points_normalized]
+    
+    points_array = np.array(points_absolute).reshape(1, -1, 2)
+    key_points = sv.KeyPoints(xy=points_array)
+    vertex_annotator = sv.VertexAnnotator(radius=5, color=sv.Color.RED)
+    annotated_image = vertex_annotator.annotate(scene=annotated_image, key_points=key_points)
+    
+    return Image.fromarray(annotated_image)
+
+@spaces.GPU
+def generate_detection_and_pointing(image: Image.Image, prompt: str, task_type: str):
+    """Main function for the detection/pointing tab."""
+    if image is None: raise gr.Error("Please upload an image.")
+    if not prompt or not prompt.strip(): raise gr.Error("Please enter a prompt describing the object.")
+
+    if task_type == "Object Detection":
+        instruction = f"You are a helpful detection assistant. Find all instances of '{prompt}' in the image. Provide their bounding box coordinates in the format [[x_min, y_min, x_max, y_max]]. The coordinates must be normalized between 0 and 1. Only output the list of coordinates."
+    else: # Point Detection
+        instruction = f"You are a helpful detection assistant. Find the center point of all instances of '{prompt}' in the image. Provide their coordinates in the format [[x, y]]. The coordinates must be normalized between 0 and 1. Only output the list of coordinates."
+
+    messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": instruction}]}]
+    prompt_full = processor_q3vl.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = processor_q3vl(text=[prompt_full], images=[image], return_tensors="pt", padding=True).to(device)
+    
+    output_ids = model_q3vl.generate(**inputs, max_new_tokens=128, do_sample=False)
+    response_text = processor_q3vl.batch_decode(output_ids, skip_special_tokens=True)[0]
+    
+    parts = response_text.split("ASSISTANT:")
+    response_text = parts[-1].strip() if len(parts) > 1 else response_text.split("<|im_end|>")[-1].strip()
+
+    coords = parse_model_output_for_coords(response_text, task_type)
+    
+    annotated_image = image
+    if coords:
+        try:
+            if task_type == "Object Detection":
+                annotated_image = create_annotated_image_from_normalized(image, coords, prompt)
+            else:
+                annotated_image = create_point_annotated_image_from_normalized(image, coords)
+        except Exception as e:
+            response_text += f"\n\n[Error] Failed to draw annotations. Details: {e}"
+            annotated_image = image
+    else:
+        response_text += "\n\n[Info] Could not find or parse coordinates from model output. No annotations were drawn."
+
+    return annotated_image, response_text
+
+# --- Gradio UI ---
 
-# Define examples for image and video inference
 image_examples = [
-    ["Describe the safety measures in the image. Conclude (Safe / Unsafe)..", "images/5.jpg"],
-    ["Convert this page to doc [markdown] precisely.", "images/3.png"],
-    ["Convert this page to doc [markdown] precisely.", "images/4.png"],
-    ["Explain the creativity in the image.", "images/6.jpg"],
-    ["Convert this page to doc [markdown] precisely.", "images/1.png"],
-    ["Convert chart to OTSL.", "images/2.png"]
+    ["Describe the safety measures in the image. Conclude (Safe / Unsafe)..", "https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/images/5.jpg"],
+    ["Convert this page to doc [markdown] precisely.", "https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/images/3.png"],
+    ["Explain the creativity in the image.", "https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/images/6.jpg"],
 ]
 
 video_examples = [
-    ["Explain the video in detail.", "videos/2.mp4"],
-    ["Explain the ad in detail.", "videos/1.mp4"]
+    ["Explain the video in detail.", "https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/videos/2.mp4"],
+    ["Explain the ad in detail.", "https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/videos/1.mp4"]
 ]
 
+detection_examples = [
+    ["https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/images/6.jpg", "Object Detection", "the person"],
+    ["https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/resolve/main/images/5.jpg", "Point Detection", "the fire extinguisher"],
+]
+
+
 css = """
 .submit-btn { background-color: #2980b9 !important; color: white !important; }
 .submit-btn:hover { background-color: #3498db !important; }
 .canvas-output { border: 2px solid #4682B4; border-radius: 10px; padding: 20px; }
 """
 
-# Create the Gradio Interface
 with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
     gr.Markdown("# **[Multimodal VLM Thinking with Qwen3-VL](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct)**")
-    with gr.Row():
-        with gr.Column():
-            with gr.Tabs():
-                with gr.TabItem("Image Inference"):
+    
+    with gr.Tabs():
+        with gr.TabItem("Image Inference"):
+            with gr.Row():
+                with gr.Column():
                     image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                     image_upload = gr.Image(type="pil", label="Image", height=290)
                     image_submit = gr.Button("Submit", elem_classes="submit-btn")
                     gr.Examples(examples=image_examples, inputs=[image_query, image_upload])
-                with gr.TabItem("Video Inference"):
+                with gr.Column():
+                     with gr.Column(elem_classes="canvas-output"):
+                        gr.Markdown("## Output")
+                        output_img, markdown_output_img = gr.Textbox(label="Raw Output Stream", interactive=False, lines=15, show_copy_button=True), gr.Markdown(label="(Result.Md)")
+
+        with gr.TabItem("Video Inference"):
+            with gr.Row():
+                with gr.Column():
                     video_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                     video_upload = gr.Video(label="Video", height=290)
                     video_submit = gr.Button("Submit", elem_classes="submit-btn")
                     gr.Examples(examples=video_examples, inputs=[video_query, video_upload])
+                with gr.Column():
+                     with gr.Column(elem_classes="canvas-output"):
+                        gr.Markdown("## Output")
+                        output_vid, markdown_output_vid = gr.Textbox(label="Raw Output Stream", interactive=False, lines=15, show_copy_button=True), gr.Markdown(label="(Result.Md)")
+
+        with gr.TabItem("Object Detection & Pointing"):
+            with gr.Row():
+                with gr.Column(scale=1):
+                    detection_image_input = gr.Image(label="Upload an image", type="pil", height=400)
+                    detection_task_type = gr.Radio(choices=["Object Detection", "Point Detection"], label="Task Type", value="Object Detection")
+                    detection_prompt_input = gr.Textbox(label="Object to Detect/Point", placeholder="e.g., 'car', 'the person's face'")
+                    detection_submit_btn = gr.Button(value="🚀 Find Objects", variant="primary")
+                with gr.Column(scale=1):
+                    detection_output_image = gr.Image(type="pil", label="Result", height=400)
+                    detection_output_textbox = gr.Textbox(label="Model Raw Output (Coordinates)", lines=10, show_copy_button=True)
+            gr.Examples(examples=detection_examples, inputs=[detection_image_input, detection_task_type, detection_prompt_input])
 
-            with gr.Accordion("Advanced options", open=False):
-                max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
-                temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
-                top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
-                top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
-                repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
-
-        with gr.Column():
-            with gr.Column(elem_classes="canvas-output"):
-                gr.Markdown("## Output")
-                output = gr.Textbox(label="Raw Output Stream", interactive=False, lines=5, show_copy_button=True)
-                with gr.Accordion("(Result.md)", open=False):
-                    markdown_output = gr.Markdown(label="(Result.Md)")
-            gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/discussions)")
-            gr.Markdown("> Using **[Qwen/Qwen3-VL-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct)**, a powerful and versatile vision-language model. It excels at understanding and processing both text and visual information, making it suitable for a wide range of multimodal tasks. The model demonstrates strong performance in areas like visual question answering, image captioning, and video analysis.")
-            gr.Markdown("> ⚠️ Note: Video inference performance can vary depending on the complexity and length of the video.")
+    with gr.Accordion("Advanced options", open=False):
+        max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
+        temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
+        top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
+        top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
+        repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
+        
+    gr.Markdown("---")
+    gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/discussions)")
+    gr.Markdown("> Using **[Qwen/Qwen3-VL-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct)**, a powerful and versatile vision-language model. It excels at understanding and processing both text and visual information, making it suitable for a wide range of multimodal tasks.")
+    gr.Markdown("> ⚠️ Note: Video inference performance can vary depending on the complexity and length of the video.")
 
+    # Event Handlers
     image_submit.click(
         fn=generate_image,
         inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
-        outputs=[output, markdown_output]
+        outputs=[output_img, markdown_output_img]
     )
     video_submit.click(
         fn=generate_video,
         inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
-        outputs=[output, markdown_output]
+        outputs=[output_vid, markdown_output_vid]
+    )
+    detection_submit_btn.click(
+        fn=generate_detection_and_pointing,
+        inputs=[detection_image_input, detection_prompt_input, detection_task_type],
+        outputs=[detection_output_image, detection_output_textbox]
     )
 
 if __name__ == "__main__":