Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -5,57 +5,69 @@ import torch
|
|
| 5 |
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
| 6 |
from qwen_vl_utils import process_vision_info
|
| 7 |
from peft import PeftModel
|
| 8 |
-
from pathlib import Path
|
| 9 |
|
| 10 |
system_prompt = (
|
| 11 |
"A conversation between User and Assistant. The user asks a question, and the Assistant solves it. "
|
| 12 |
-
"
|
| 13 |
"El razonamiento y la respuesta van en <think></think> y <answer></answer>."
|
| 14 |
)
|
| 15 |
|
| 16 |
-
MODEL_ID
|
| 17 |
-
ADAPTER_ID =
|
| 18 |
|
| 19 |
processor = AutoProcessor.from_pretrained(MODEL_ID)
|
| 20 |
|
| 21 |
-
|
| 22 |
-
has_gpu = torch.cuda.is_available()
|
| 23 |
attn_impl = "flash_attention_2" if has_gpu else "eager"
|
| 24 |
-
model
|
| 25 |
MODEL_ID,
|
| 26 |
torch_dtype=torch.bfloat16,
|
| 27 |
attn_implementation=attn_impl,
|
| 28 |
-
device_map="auto"
|
| 29 |
)
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
model = PeftModel.from_pretrained(model, ADAPTER_ID)
|
| 33 |
-
model = model.merge_and_unload()
|
| 34 |
-
model.eval()
|
| 35 |
-
device = torch.device("cuda" if has_gpu else "cpu")
|
| 36 |
-
model.to(device)
|
| 37 |
|
| 38 |
example_imgs = [
|
| 39 |
-
("6.png",
|
| 40 |
("163.png", "Tienda esquinera"),
|
| 41 |
("img_71_2.png", "Comida colombiana"),
|
| 42 |
-
("img_98.png",
|
| 43 |
]
|
| 44 |
|
| 45 |
-
def cargar_imagen(
|
| 46 |
-
return Image.open(
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
gr.Markdown(
|
| 60 |
"""
|
| 61 |
<h1>🇨🇴
|
|
@@ -65,153 +77,111 @@ with gr.Blocks(theme='lone17/kotaemon') as demo:
|
|
| 65 |
"""
|
| 66 |
)
|
| 67 |
|
|
|
|
| 68 |
with gr.Row():
|
| 69 |
-
|
| 70 |
-
with gr.Column(scale=1):
|
| 71 |
gr.Markdown(
|
| 72 |
"""
|
| 73 |
#### 📌 Motivación del proyecto
|
| 74 |
-
|
| 75 |
<br/>
|
| 76 |
-
|
| 77 |
#### 🌟 Impacto
|
| 78 |
-
|
| 79 |
|
| 80 |
#### 👥 Equipo
|
| 81 |
• Fabian Perez
|
| 82 |
• Henry Mantilla
|
| 83 |
• Andrea Parra
|
| 84 |
• Juan Calderón
|
| 85 |
-
•
|
| 86 |
"""
|
| 87 |
)
|
| 88 |
-
|
| 89 |
-
# Columna derecha
|
| 90 |
-
with gr.Column(scale=1):
|
| 91 |
gr.Markdown(
|
| 92 |
"""
|
| 93 |
#### 🚀 Ideas futuras
|
| 94 |
-
- 📈 Escalar
|
| 95 |
-
- 🎤
|
| 96 |
-
- 🌐 Traducción automática
|
| 97 |
-
- 🗺️ Más dialectos
|
| 98 |
-
- 🔄 Retroalimentación
|
| 99 |
-
- 🗺️ Mapas
|
| 100 |
-
|
| 101 |
#### 🤖 Modelos utilizados
|
| 102 |
- *Qwen2.5-VL-3B-Instruct*
|
| 103 |
- Dataset: [QuestionAnswer-ImgsColombia](https://huggingface.co/datasets/4nd/QuestionAnswer-ImgsColombia)
|
| 104 |
"""
|
| 105 |
)
|
| 106 |
|
| 107 |
-
|
| 108 |
-
|
| 109 |
with gr.Row(equal_height=True):
|
| 110 |
-
# Columna izquierda
|
| 111 |
with gr.Column(scale=1):
|
| 112 |
pregunta = gr.Textbox(
|
| 113 |
label="❓ Pregunta sobre tu imagen",
|
| 114 |
placeholder="¿Qué muestra esta imagen?",
|
| 115 |
-
lines=2
|
| 116 |
)
|
| 117 |
|
| 118 |
-
# Asignamos elem_id al Gallery directamente
|
| 119 |
galeria = gr.Gallery(
|
| 120 |
label="📁 Elige una imagen de ejemplo",
|
| 121 |
value=[img for img, _ in example_imgs],
|
| 122 |
-
columns=
|
| 123 |
-
height=
|
| 124 |
allow_preview=True,
|
| 125 |
show_label=True,
|
| 126 |
-
elem_id="galeria-scroll"
|
| 127 |
)
|
| 128 |
|
| 129 |
-
# Columna derecha
|
| 130 |
with gr.Column(scale=1):
|
| 131 |
imagen_mostrada = gr.Image(
|
| 132 |
label="🖼 Imagen seleccionada o subida",
|
| 133 |
type="numpy",
|
| 134 |
-
height=256
|
| 135 |
)
|
| 136 |
-
|
| 137 |
respuesta = gr.Textbox(
|
| 138 |
label="🧠 Respuesta",
|
| 139 |
interactive=False,
|
| 140 |
-
lines=4
|
| 141 |
)
|
| 142 |
-
|
| 143 |
btn_procesar = gr.Button("🔍 Procesar")
|
| 144 |
|
| 145 |
def seleccionar_imagen(evt: gr.SelectData):
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
pil = cargar_imagen(img_path)
|
| 149 |
-
return np.array(pil)
|
| 150 |
-
|
| 151 |
galeria.select(fn=seleccionar_imagen, inputs=None, outputs=imagen_mostrada)
|
| 152 |
|
| 153 |
def responder(img, pregunta_text):
|
| 154 |
if img is None or pregunta_text.strip() == "":
|
| 155 |
return "Por favor sube una imagen y escribe una pregunta."
|
| 156 |
|
| 157 |
-
# Convertir array numpy a PIL si es necesario
|
| 158 |
if isinstance(img, np.ndarray):
|
| 159 |
-
|
| 160 |
-
else:
|
| 161 |
-
img_pil = img # ya es PIL
|
| 162 |
|
| 163 |
messages = [
|
| 164 |
-
{
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
"content": [
|
| 171 |
-
{"type": "text", "text": pregunta_text},
|
| 172 |
-
{"type": "image", "image": img_pil},
|
| 173 |
-
],
|
| 174 |
-
}
|
| 175 |
]
|
| 176 |
-
|
| 177 |
-
text = processor.apply_chat_template(
|
| 178 |
-
messages,
|
| 179 |
-
tokenize=False,
|
| 180 |
-
add_generation_prompt=True
|
| 181 |
-
)
|
| 182 |
image_inputs, video_inputs = process_vision_info(messages)
|
| 183 |
-
|
| 184 |
inputs = processor(
|
| 185 |
text=[text],
|
| 186 |
images=image_inputs,
|
| 187 |
videos=video_inputs,
|
| 188 |
padding=True,
|
| 189 |
return_tensors="pt",
|
| 190 |
-
)
|
| 191 |
-
inputs = inputs.to(device)
|
| 192 |
|
| 193 |
with torch.no_grad():
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
top_p=1.0,
|
| 198 |
-
do_sample=True,
|
| 199 |
-
temperature=0.9
|
| 200 |
-
)
|
| 201 |
|
| 202 |
-
|
| 203 |
-
out_ids[len(in_ids):]
|
| 204 |
-
for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
| 205 |
-
]
|
| 206 |
-
respuesta_text = processor.batch_decode(
|
| 207 |
-
trimmed,
|
| 208 |
-
skip_special_tokens=True,
|
| 209 |
-
clean_up_tokenization_spaces=False
|
| 210 |
-
)
|
| 211 |
-
|
| 212 |
-
return respuesta_text[0]
|
| 213 |
-
|
| 214 |
-
btn_procesar.click(fn=responder, inputs=[imagen_mostrada, pregunta], outputs=respuesta)
|
| 215 |
|
| 216 |
if __name__ == "__main__":
|
| 217 |
demo.launch()
|
|
|
|
|
|
| 5 |
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
| 6 |
from qwen_vl_utils import process_vision_info
|
| 7 |
from peft import PeftModel
|
|
|
|
| 8 |
|
| 9 |
system_prompt = (
|
| 10 |
"A conversation between User and Assistant. The user asks a question, and the Assistant solves it. "
|
| 11 |
+
"El assistant es un experto sobre Colombia. Primero razona en mente y luego da la respuesta. "
|
| 12 |
"El razonamiento y la respuesta van en <think></think> y <answer></answer>."
|
| 13 |
)
|
| 14 |
|
| 15 |
+
MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
|
| 16 |
+
ADAPTER_ID = "Factral/qwen2.5vl-3b-colombia-finetuned"
|
| 17 |
|
| 18 |
processor = AutoProcessor.from_pretrained(MODEL_ID)
|
| 19 |
|
| 20 |
+
has_gpu = torch.cuda.is_available()
|
|
|
|
| 21 |
attn_impl = "flash_attention_2" if has_gpu else "eager"
|
| 22 |
+
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
| 23 |
MODEL_ID,
|
| 24 |
torch_dtype=torch.bfloat16,
|
| 25 |
attn_implementation=attn_impl,
|
| 26 |
+
device_map="auto",
|
| 27 |
)
|
| 28 |
+
model = PeftModel.from_pretrained(model, ADAPTER_ID).merge_and_unload()
|
| 29 |
+
model.eval().to(torch.device("cuda" if has_gpu else "cpu"))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
| 31 |
example_imgs = [
|
| 32 |
+
("6.png", "Shakira"),
|
| 33 |
("163.png", "Tienda esquinera"),
|
| 34 |
("img_71_2.png", "Comida colombiana"),
|
| 35 |
+
("img_98.png", "Oso de anteojos"),
|
| 36 |
]
|
| 37 |
|
| 38 |
+
def cargar_imagen(path: str) -> Image.Image:
|
| 39 |
+
return Image.open(path)
|
| 40 |
+
|
| 41 |
+
CSS_CUSTOM = """
|
| 42 |
+
/* Galería horizontal con miniaturas */
|
| 43 |
+
#galeria-scroll {
|
| 44 |
+
overflow-x: auto;
|
| 45 |
+
overflow-y: hidden;
|
| 46 |
+
padding: 4px;
|
| 47 |
+
scrollbar-width: thin;
|
| 48 |
+
}
|
| 49 |
+
#galeria-scroll .gallery { flex-wrap: nowrap !important; }
|
| 50 |
+
#galeria-scroll .gallery-item {
|
| 51 |
+
flex: 0 0 auto !important;
|
| 52 |
+
width: 90px !important;
|
| 53 |
+
height: 90px !important;
|
| 54 |
+
margin-right: 6px;
|
| 55 |
+
}
|
| 56 |
+
#galeria-scroll .gallery-item img { object-fit: cover; }
|
| 57 |
+
|
| 58 |
+
/* Texto blanco y sin halo azul al enfocar */
|
| 59 |
+
input, textarea { color: #fff !important; }
|
| 60 |
+
input::placeholder, textarea::placeholder { color: #ddd !important; }
|
| 61 |
+
label { color: #fff !important; }
|
| 62 |
+
input:focus, textarea:focus {
|
| 63 |
+
outline: none !important;
|
| 64 |
+
box-shadow: none !important;
|
| 65 |
+
border-color: #888 !important;
|
| 66 |
+
}
|
| 67 |
+
"""
|
| 68 |
+
|
| 69 |
+
with gr.Blocks(theme="lone17/kotaemon", css=CSS_CUSTOM) as demo:
|
| 70 |
+
# título
|
| 71 |
gr.Markdown(
|
| 72 |
"""
|
| 73 |
<h1>🇨🇴
|
|
|
|
| 77 |
"""
|
| 78 |
)
|
| 79 |
|
| 80 |
+
# motivación / ideas futuras en dos columnas
|
| 81 |
with gr.Row():
|
| 82 |
+
with gr.Column():
|
|
|
|
| 83 |
gr.Markdown(
|
| 84 |
"""
|
| 85 |
#### 📌 Motivación del proyecto
|
| 86 |
+
BacanoResponder permite a los usuarios colombianos obtener información contextual de sus imágenes.
|
| 87 |
<br/>
|
| 88 |
+
|
| 89 |
#### 🌟 Impacto
|
| 90 |
+
Difunde cultura local y apoya a estudiantes, turistas y creadores de contenido.
|
| 91 |
|
| 92 |
#### 👥 Equipo
|
| 93 |
• Fabian Perez
|
| 94 |
• Henry Mantilla
|
| 95 |
• Andrea Parra
|
| 96 |
• Juan Calderón
|
| 97 |
+
• [SemilleroCV](https://semillerocv.github.io/)
|
| 98 |
"""
|
| 99 |
)
|
| 100 |
+
with gr.Column():
|
|
|
|
|
|
|
| 101 |
gr.Markdown(
|
| 102 |
"""
|
| 103 |
#### 🚀 Ideas futuras
|
| 104 |
+
- 📈 Escalar el dataset
|
| 105 |
+
- 🎤 Soporte de voz en dialectos regionales
|
| 106 |
+
- 🌐 Traducción automática
|
| 107 |
+
- 🗺️ Más dialectos/costumbres
|
| 108 |
+
- 🔄 Retroalimentación continua
|
| 109 |
+
- 🗺️ Mapas turísticos
|
| 110 |
+
|
| 111 |
#### 🤖 Modelos utilizados
|
| 112 |
- *Qwen2.5-VL-3B-Instruct*
|
| 113 |
- Dataset: [QuestionAnswer-ImgsColombia](https://huggingface.co/datasets/4nd/QuestionAnswer-ImgsColombia)
|
| 114 |
"""
|
| 115 |
)
|
| 116 |
|
|
|
|
|
|
|
| 117 |
with gr.Row(equal_height=True):
|
|
|
|
| 118 |
with gr.Column(scale=1):
|
| 119 |
pregunta = gr.Textbox(
|
| 120 |
label="❓ Pregunta sobre tu imagen",
|
| 121 |
placeholder="¿Qué muestra esta imagen?",
|
| 122 |
+
lines=2,
|
| 123 |
)
|
| 124 |
|
|
|
|
| 125 |
galeria = gr.Gallery(
|
| 126 |
label="📁 Elige una imagen de ejemplo",
|
| 127 |
value=[img for img, _ in example_imgs],
|
| 128 |
+
columns=999, # fuerza una sola fila
|
| 129 |
+
height="110px",
|
| 130 |
allow_preview=True,
|
| 131 |
show_label=True,
|
| 132 |
+
elem_id="galeria-scroll",
|
| 133 |
)
|
| 134 |
|
|
|
|
| 135 |
with gr.Column(scale=1):
|
| 136 |
imagen_mostrada = gr.Image(
|
| 137 |
label="🖼 Imagen seleccionada o subida",
|
| 138 |
type="numpy",
|
| 139 |
+
height=256,
|
| 140 |
)
|
|
|
|
| 141 |
respuesta = gr.Textbox(
|
| 142 |
label="🧠 Respuesta",
|
| 143 |
interactive=False,
|
| 144 |
+
lines=4,
|
| 145 |
)
|
|
|
|
| 146 |
btn_procesar = gr.Button("🔍 Procesar")
|
| 147 |
|
| 148 |
def seleccionar_imagen(evt: gr.SelectData):
|
| 149 |
+
path = example_imgs[evt.index][0]
|
| 150 |
+
return np.array(cargar_imagen(path))
|
|
|
|
|
|
|
|
|
|
| 151 |
galeria.select(fn=seleccionar_imagen, inputs=None, outputs=imagen_mostrada)
|
| 152 |
|
| 153 |
def responder(img, pregunta_text):
|
| 154 |
if img is None or pregunta_text.strip() == "":
|
| 155 |
return "Por favor sube una imagen y escribe una pregunta."
|
| 156 |
|
|
|
|
| 157 |
if isinstance(img, np.ndarray):
|
| 158 |
+
img = Image.fromarray(img.astype("uint8"))
|
|
|
|
|
|
|
| 159 |
|
| 160 |
messages = [
|
| 161 |
+
{"role": "system", "content": [{"type": "text", "text": system_prompt}]},
|
| 162 |
+
{"role": "user",
|
| 163 |
+
"content": [
|
| 164 |
+
{"type": "text", "text": pregunta_text},
|
| 165 |
+
{"type": "image", "image": img},
|
| 166 |
+
]},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
]
|
| 168 |
+
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
image_inputs, video_inputs = process_vision_info(messages)
|
|
|
|
| 170 |
inputs = processor(
|
| 171 |
text=[text],
|
| 172 |
images=image_inputs,
|
| 173 |
videos=video_inputs,
|
| 174 |
padding=True,
|
| 175 |
return_tensors="pt",
|
| 176 |
+
).to(model.device)
|
|
|
|
| 177 |
|
| 178 |
with torch.no_grad():
|
| 179 |
+
out_ids = model.generate(**inputs, max_new_tokens=512, top_p=1.0, do_sample=True, temperature=0.9)
|
| 180 |
+
trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, out_ids)]
|
| 181 |
+
return processor.batch_decode(trimmed, skip_special_tokens=True)[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
| 182 |
|
| 183 |
+
btn_procesar.click(responder, inputs=[imagen_mostrada, pregunta], outputs=respuesta)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 184 |
|
| 185 |
if __name__ == "__main__":
|
| 186 |
demo.launch()
|
| 187 |
+
|