Update app.py

app.py

@@ -2,20 +2,21 @@ import gradio as gr
 from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 from PIL import Image
+import re  # Importando o módulo de expressões regulares
 import requests
 from io import BytesIO
 
-# Load the Qwen-VL model and tokenizer
+# Carregar o modelo Qwen-VL e o tokenizer
 tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-VL", trust_remote_code=True)
 model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL", device_map="cuda", trust_remote_code=True).eval()
 
 def generate_predictions(image_input, text_input):
-    # Save the image locally to match the original example
+    # Inverter a imagem para corrigir o negativo
     user_image_path = "/tmp/user_input_test_image.jpg"
-    image_input.save(user_image_path)
-    image_input = Image.open(user_image_path)
-
-    # Prepare the inputs
+    Image.fromarray((255 - (image_input * 255).astype('uint8'))).save(user_image_path)
+    
+    
+    # Preparar as entradas
     query = tokenizer.from_list_format([
         {'image': user_image_path},
         {'text': text_input},
@@ -23,20 +24,45 @@ def generate_predictions(image_input, text_input):
     inputs = tokenizer(query, return_tensors='pt')
     inputs = inputs.to(model.device)
     
-    # Generate the caption
+    # Gerar a legenda
     pred = model.generate(**inputs)
-    response = tokenizer.decode(pred.cpu()[0], skip_special_tokens=False)
+    full_response = tokenizer.decode(pred.cpu()[0], skip_special_tokens=False)
+    
+    # Remover o texto de input e outras partes indesejadas da resposta completa
+    frontend_response = re.sub(r'Picture \d+:|<.*?>|\/tmp\/.*\.jpg', '', full_response).replace(text_input, '').strip()
     
-    # Draw bounding boxes if any
-    image_with_boxes = tokenizer.draw_bbox_on_latest_picture(response)
+    # Desenhar caixas delimitadoras, se houver
+    image_with_boxes = tokenizer.draw_bbox_on_latest_picture(full_response)
     
-    return image_with_boxes, response
+    # Salvar e recarregar a imagem para garantir que seja uma imagem PIL
+    if image_with_boxes:
+        temp_path = "/tmp/image_with_boxes.jpg"
+        image_with_boxes.save(temp_path)
+        image_with_boxes = Image.open(temp_path)
+    
+    return image_with_boxes, frontend_response  # Retornando a resposta formatada para o frontend
 
-# Create Gradio Interface
+# Criar interface Gradio
+# Create Gradio interface
 iface = gr.Interface(
     fn=generate_predictions, 
-    inputs=["image", "text"], 
-    outputs=["image", "text"]
+    inputs=[
+        gr.inputs.Image(label="Image Input"), 
+        gr.inputs.Textbox(default="Generate a caption for that image with grounding:", label="Prompt")
+    ], 
+    outputs=[
+        gr.outputs.Image(type='pil', label="Image"),  # Explicitly set type to 'pil'
+        gr.outputs.Textbox(label="Generated")
+    ],
+    title="Qwen-VL Demonstration",
+    description = """
+## Qwen-VL: A Multimodal Large Vision Language Model by Alibaba Cloud
+**Space by [@Artificialguybr](https://twitter.com/artificialguybr)**
+    
+### Key Features:
+- **Strong Performance**: Surpasses existing LVLMs on multiple English benchmarks including Zero-shot Captioning and VQA.
+- **Multi-lingual Support**: Supports English, Chinese, and multi-lingual conversation.
+- **High Resolution**: Utilizes 448*448 resolution for fine-grained recognition and understanding.
+""",
 )
-
-iface.launch()
+iface.launch(share=True)