Spaces:
Runtime error
Runtime error
| import gradio as gr | |
| from PIL import Image | |
| import numpy as np | |
| import torch | |
| from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor | |
| from qwen_vl_utils import process_vision_info | |
| from peft import PeftModel | |
| system_prompt = ( | |
| "A conversation between User and Assistant. The user asks a question, and the Assistant solves it. " | |
| "El assistant es un experto sobre Colombia. Primero razona en mente y luego da la respuesta. " | |
| "El razonamiento y la respuesta van en <think></think> y <answer></answer>." | |
| ) | |
| MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct" | |
| ADAPTER_ID = "Factral/qwen2.5vl-3b-colombia-finetuned" | |
| processor = AutoProcessor.from_pretrained(MODEL_ID) | |
| has_gpu = torch.cuda.is_available() | |
| attn_impl = "flash_attention_2" if has_gpu else "eager" | |
| model = Qwen2_5_VLForConditionalGeneration.from_pretrained( | |
| MODEL_ID, | |
| torch_dtype=torch.bfloat16, | |
| attn_implementation=attn_impl, | |
| device_map="auto", | |
| ) | |
| model = PeftModel.from_pretrained(model, ADAPTER_ID).merge_and_unload() | |
| model.eval().to(torch.device("cuda" if has_gpu else "cpu")) | |
| example_imgs = [ | |
| ("6.png", "Shakira"), | |
| ("163.png", "Tienda esquinera"), | |
| ("img_71_2.png", "Comida colombiana"), | |
| ("img_98.png", "Oso de anteojos"), | |
| ] | |
| def cargar_imagen(path: str) -> Image.Image: | |
| return Image.open(path) | |
| CSS_CUSTOM = """ | |
| /* Galería horizontal con miniaturas */ | |
| #galeria-scroll { | |
| overflow-x: auto; | |
| overflow-y: hidden; | |
| padding: 4px; | |
| scrollbar-width: thin; | |
| } | |
| #galeria-scroll .gallery { flex-wrap: nowrap !important; } | |
| #galeria-scroll .gallery-item { | |
| flex: 0 0 auto !important; | |
| width: 90px !important; | |
| height: 90px !important; | |
| margin-right: 6px; | |
| } | |
| #galeria-scroll .gallery-item img { object-fit: cover; } | |
| /* Texto blanco y sin halo azul al enfocar */ | |
| input, textarea { color: #fff !important; } | |
| input::placeholder, textarea::placeholder { color: #ddd !important; } | |
| label { color: #fff !important; } | |
| .gr-text-input:focus-within, | |
| .gr-text-area:focus-within, | |
| .gr-input:focus-within { | |
| outline: none !important; | |
| box-shadow: none !important; | |
| border-color: #888 !important; /* gris neutro opcional */ | |
| } | |
| /* Por si quedaba algo en el propio input/textarea */ | |
| input:focus, textarea:focus, | |
| input:focus-visible, textarea:focus-visible { | |
| outline: none !important; | |
| box-shadow: none !important; | |
| border-color: #888 !important; | |
| """ | |
| with gr.Blocks(theme="lone17/kotaemon", css=CSS_CUSTOM) as demo: | |
| # título | |
| gr.Markdown( | |
| """ | |
| <h1>🇨🇴 | |
| <span style='color:gold;'>Bacan</span><span style='color:blue;'>oResp</span><span style='color:red;'>onder</span> | |
| </h1> | |
| <p>Sube o elige una imagen, haz una pregunta y obtén una respuesta con contexto local.</p> | |
| """ | |
| ) | |
| # motivación / ideas futuras en dos columnas | |
| with gr.Row(): | |
| with gr.Column(): | |
| gr.Markdown( | |
| """ | |
| #### 📌 Motivación del proyecto | |
| BacanoResponder permite a los usuarios colombianos obtener información contextual de sus imágenes. | |
| <br/> | |
| #### 🌟 Impacto | |
| Difunde cultura local y apoya a estudiantes, turistas y creadores de contenido. | |
| #### 👥 Equipo | |
| • Fabian Perez | |
| • Henry Mantilla | |
| • Andrea Parra | |
| • Juan Calderón | |
| • Semillero de Investigación del que somos parte [SemilleroCV](https://semillerocv.github.io/) | |
| """ | |
| ) | |
| with gr.Column(): | |
| gr.Markdown( | |
| """ | |
| #### 🚀 Ideas futuras | |
| - 📈 Escalar el dataset | |
| - 🎤 Soporte de voz en dialectos regionales | |
| - 🌐 Traducción automática | |
| - 🗺️ Más dialectos/costumbres | |
| - 🔄 Retroalimentación continua | |
| - 🗺️ Mapas turísticos | |
| #### 🤖 Modelos utilizados | |
| - *Qwen2.5-VL-3B-Instruct* | |
| - Dataset: [QuestionAnswer-ImgsColombia](https://huggingface.co/datasets/4nd/QuestionAnswer-ImgsColombia) | |
| """ | |
| ) | |
| with gr.Row(equal_height=True): | |
| with gr.Column(scale=1): | |
| pregunta = gr.Textbox( | |
| label="❓ Pregunta sobre tu imagen", | |
| placeholder="¿Qué muestra esta imagen?", | |
| lines=2, | |
| ) | |
| galeria = gr.Gallery( | |
| label="📁 Elige una imagen de ejemplo", | |
| value=[img for img, _ in example_imgs], | |
| columns=3, | |
| height="384px", | |
| allow_preview=True, | |
| show_label=True, | |
| elem_id="galeria-scroll", | |
| ) | |
| with gr.Column(scale=1): | |
| imagen_mostrada = gr.Image( | |
| label="🖼 Imagen seleccionada o subida", | |
| type="numpy", | |
| height=256, | |
| ) | |
| respuesta = gr.Textbox( | |
| label="🧠 Respuesta", | |
| interactive=False, | |
| lines=4, | |
| ) | |
| btn_procesar = gr.Button("🔍 Procesar") | |
| def seleccionar_imagen(evt: gr.SelectData): | |
| path = example_imgs[evt.index][0] | |
| return np.array(cargar_imagen(path)) | |
| galeria.select(fn=seleccionar_imagen, inputs=None, outputs=imagen_mostrada) | |
| def responder(img, pregunta_text): | |
| if img is None or pregunta_text.strip() == "": | |
| return "Por favor sube una imagen y escribe una pregunta." | |
| if isinstance(img, np.ndarray): | |
| img = Image.fromarray(img.astype("uint8")) | |
| messages = [ | |
| {"role": "system", "content": [{"type": "text", "text": system_prompt}]}, | |
| {"role": "user", | |
| "content": [ | |
| {"type": "text", "text": pregunta_text}, | |
| {"type": "image", "image": img}, | |
| ]}, | |
| ] | |
| 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", | |
| ).to(model.device) | |
| with torch.no_grad(): | |
| out_ids = model.generate(**inputs, max_new_tokens=512, top_p=1.0, do_sample=True, temperature=0.9) | |
| trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, out_ids)] | |
| return processor.batch_decode(trimmed, skip_special_tokens=True)[0] | |
| btn_procesar.click(responder, inputs=[imagen_mostrada, pregunta], outputs=respuesta) | |
| if __name__ == "__main__": | |
| demo.launch() | |