Update app.py

app.py

@@ -2,21 +2,20 @@ import gradio as gr
 from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 from PIL import Image
-import re  # Importando o módulo de expressões regulares
+import re 
 import requests
 from io import BytesIO
 
-# Carregar o modelo Qwen-VL e o tokenizer
 tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-VL-Chat-Int4", trust_remote_code=True)
 model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-VL-Chat-Int4", device_map="auto", trust_remote_code=True).eval()
 
-def generate_predictions(image_input, text_input):
-    # Inverter a imagem para corrigir o negativo
+def generate_predictions(image_input, text_input, with_grounding):
     user_image_path = "/tmp/user_input_test_image.jpg"
     Image.fromarray((255 - (image_input * 255).astype('uint8'))).save(user_image_path)
     
+    if with_grounding == "Yes":
+        text_input += " with grounding"
     
-    # Preparar as entradas
     query = tokenizer.from_list_format([
         {'image': user_image_path},
         {'text': text_input},
@@ -24,34 +23,30 @@ def generate_predictions(image_input, text_input):
     inputs = tokenizer(query, return_tensors='pt')
     inputs = inputs.to(model.device)
     
-    # Gerar a legenda
     pred = model.generate(**inputs)
     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()
+    print("Generated Caption:", frontend_response)  # Debugging line
     
-    # Desenhar caixas delimitadoras, se houver
     image_with_boxes = tokenizer.draw_bbox_on_latest_picture(full_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
+    return image_with_boxes, frontend_response
 
-# Criar interface Gradio
-# Create Gradio interface
 iface = gr.Interface(
     fn=generate_predictions, 
     inputs=[
         gr.inputs.Image(label="Image Input"), 
-        gr.inputs.Textbox(default="Generate a caption for that image with grounding:", label="Prompt")
+        gr.inputs.Textbox(default="Generate a caption for that image:", label="Prompt"),
+        gr.inputs.Radio(["No", "Yes"], label="With Grounding", default="No")
     ], 
     outputs=[
-        gr.outputs.Image(type='pil', label="Image"),  # Explicitly set type to 'pil'
+        gr.outputs.Image(type='pil', label="Image"),
         gr.outputs.Textbox(label="Generated")
     ],
     title="Qwen-VL Demonstration",
@@ -65,4 +60,4 @@ iface = gr.Interface(
 - **High Resolution**: Utilizes 448*448 resolution for fine-grained recognition and understanding.
 """,
 )
-iface.launch()
+iface.launch()