Spaces:
Running on Zero
Running on Zero
Upload app.py
Browse files
app.py
CHANGED
|
@@ -1,239 +1,177 @@
|
|
| 1 |
import os
|
| 2 |
-
import re
|
| 3 |
-
from pathlib import Path
|
| 4 |
-
|
| 5 |
import gradio as gr
|
| 6 |
import numpy as np
|
| 7 |
import torch
|
|
|
|
| 8 |
from huggingface_hub import hf_hub_download
|
| 9 |
from PIL import Image
|
| 10 |
-
from transformers import CLIPModel, CLIPProcessor
|
| 11 |
import spaces # <-- Importante para Hugging Face ZeroGPU
|
| 12 |
|
| 13 |
-
# --- IMPORTACIONES DE
|
|
|
|
| 14 |
from sam2.build_sam import build_sam2
|
| 15 |
-
from sam2.
|
| 16 |
|
| 17 |
-
# --- CONFIGURACIÓN DE
|
|
|
|
| 18 |
SAM2_REPO = "facebook/sam2.1-hiera-base-plus"
|
| 19 |
CHECKPOINT_FILENAME = "sam2.1_hiera_base_plus.pt"
|
| 20 |
SAM2_CONFIG = "configs/sam2.1/sam2.1_hiera_b+.yaml"
|
| 21 |
|
| 22 |
-
|
|
|
|
| 23 |
|
| 24 |
-
# Volvemos a CUDA, ya que ahora cargaremos los modelos dentro de la función autorizada
|
| 25 |
DEVICE = "cuda"
|
| 26 |
-
CLIP_THRESHOLD = 0.26
|
| 27 |
|
| 28 |
-
# Variables globales para
|
| 29 |
-
|
| 30 |
-
|
| 31 |
clip_model = None
|
| 32 |
clip_processor = None
|
| 33 |
|
| 34 |
COLOR_PALETTE = [
|
| 35 |
-
(
|
| 36 |
-
(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
]
|
| 38 |
|
| 39 |
-
def
|
| 40 |
-
"""Descarga el modelo SAM 2.1 desde Hugging Face."""
|
| 41 |
cache_dir = Path("./models")
|
| 42 |
cache_dir.mkdir(parents=True, exist_ok=True)
|
| 43 |
local_path = cache_dir / CHECKPOINT_FILENAME
|
| 44 |
-
|
| 45 |
if not local_path.exists():
|
| 46 |
-
print(f"Descargando {CHECKPOINT_FILENAME}
|
| 47 |
-
local_path = Path(
|
| 48 |
-
hf_hub_download(
|
| 49 |
-
repo_id=SAM2_REPO,
|
| 50 |
-
filename=CHECKPOINT_FILENAME,
|
| 51 |
-
cache_dir=str(cache_dir),
|
| 52 |
-
)
|
| 53 |
-
)
|
| 54 |
-
print("¡Descarga completada!")
|
| 55 |
return str(local_path)
|
| 56 |
|
| 57 |
-
def create_mask_overlay(image: Image.Image,
|
| 58 |
-
|
| 59 |
-
overlay_image = image.copy()
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
sorted_masks = sorted(masks, key=(lambda x: x["area"]), reverse=True)
|
| 63 |
-
|
| 64 |
-
for i, mask_data in enumerate(sorted_masks):
|
| 65 |
color = COLOR_PALETTE[i % len(COLOR_PALETTE)]
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
# Convertir la matriz booleana a una imagen de escala de grises (L)
|
| 69 |
-
mask_image = Image.fromarray(mask_bool.astype(np.uint8) * 255, mode="L")
|
| 70 |
-
|
| 71 |
-
# Crear una capa del mismo tamaño con el color correspondiente
|
| 72 |
-
color_overlay = Image.new("RGBA", image.size, color)
|
| 73 |
-
|
| 74 |
-
# Pegar el color transparente SOBRE la imagen original, usando la máscara
|
| 75 |
overlay_image.paste(color_overlay, (0, 0), mask_image)
|
| 76 |
|
| 77 |
return overlay_image
|
| 78 |
|
| 79 |
-
def mask_to_bbox(mask: np.ndarray):
|
| 80 |
-
ys, xs = np.where(mask.astype(np.uint8))
|
| 81 |
-
if ys.size == 0 or xs.size == 0:
|
| 82 |
-
return None
|
| 83 |
-
return int(xs.min()), int(ys.min()), int(xs.max()) + 1, int(ys.max()) + 1
|
| 84 |
-
|
| 85 |
-
def crop_masked_region(image: Image.Image, mask: np.ndarray) -> Image.Image | None:
|
| 86 |
-
bbox = mask_to_bbox(mask)
|
| 87 |
-
if bbox is None:
|
| 88 |
-
return None
|
| 89 |
-
mask_img = Image.fromarray((mask.astype(np.uint8) * 255).astype(np.uint8), mode="L")
|
| 90 |
-
background = Image.new("RGB", image.size, (127, 127, 127))
|
| 91 |
-
masked = Image.composite(image, background, mask_img)
|
| 92 |
-
return masked.crop(bbox)
|
| 93 |
-
|
| 94 |
-
def normalize_features(features: torch.Tensor | object) -> torch.Tensor:
|
| 95 |
-
if hasattr(features, "pooler_output"):
|
| 96 |
-
features = features.pooler_output
|
| 97 |
-
elif hasattr(features, "last_hidden_state"):
|
| 98 |
-
features = features.last_hidden_state[:, 0, :]
|
| 99 |
-
if not isinstance(features, torch.Tensor):
|
| 100 |
-
raise RuntimeError("No se pudieron obtener características de CLIP.")
|
| 101 |
-
return features / features.norm(dim=-1, keepdim=True)
|
| 102 |
-
|
| 103 |
-
def compute_clip_features(images: list[Image.Image]):
|
| 104 |
-
inputs = clip_processor(images=images, return_tensors="pt", padding=True).to(DEVICE)
|
| 105 |
-
with torch.no_grad():
|
| 106 |
-
features = clip_model.get_image_features(**inputs)
|
| 107 |
-
return normalize_features(features)
|
| 108 |
-
|
| 109 |
-
def select_masks_by_text(image: Image.Image, masks: list[dict], prompt: str) -> tuple[list[dict], list[tuple[str, float | None]]]:
|
| 110 |
-
terms = [t.strip() for t in re.split(r"[,\n]+", prompt) if t.strip()]
|
| 111 |
-
if len(terms) == 0:
|
| 112 |
-
return [], []
|
| 113 |
-
|
| 114 |
-
crops = []
|
| 115 |
-
valid_masks = []
|
| 116 |
-
for mask in masks:
|
| 117 |
-
crop = crop_masked_region(image, mask["segmentation"])
|
| 118 |
-
if crop is not None:
|
| 119 |
-
valid_masks.append(mask)
|
| 120 |
-
crops.append(crop)
|
| 121 |
-
|
| 122 |
-
if len(crops) == 0:
|
| 123 |
-
return [], [(term, None) for term in terms]
|
| 124 |
-
|
| 125 |
-
image_features = compute_clip_features(crops)
|
| 126 |
-
text_prompts = [f"A photo of a {term}." for term in terms]
|
| 127 |
-
text_inputs = clip_processor(text=text_prompts, return_tensors="pt", padding=True).to(DEVICE)
|
| 128 |
-
|
| 129 |
-
with torch.no_grad():
|
| 130 |
-
text_features = clip_model.get_text_features(**text_inputs)
|
| 131 |
-
text_features = normalize_features(text_features)
|
| 132 |
-
|
| 133 |
-
similarities = (image_features @ text_features.T).cpu()
|
| 134 |
-
selected_indices = set()
|
| 135 |
-
hits = []
|
| 136 |
-
|
| 137 |
-
for term_idx, term in enumerate(terms):
|
| 138 |
-
scores = similarities[:, term_idx]
|
| 139 |
-
valid_idxs = torch.where(scores >= CLIP_THRESHOLD)[0].tolist()
|
| 140 |
-
|
| 141 |
-
if valid_idxs:
|
| 142 |
-
selected_indices.update(valid_idxs)
|
| 143 |
-
best_score = float(torch.max(scores[valid_idxs]).item())
|
| 144 |
-
hits.append((term, best_score))
|
| 145 |
-
else:
|
| 146 |
-
hits.append((term, None))
|
| 147 |
-
|
| 148 |
-
selected = [valid_masks[i] for i in sorted(selected_indices)]
|
| 149 |
-
return selected, hits
|
| 150 |
-
|
| 151 |
@spaces.GPU
|
| 152 |
@torch.no_grad()
|
| 153 |
-
def
|
| 154 |
-
|
| 155 |
-
global sam2_model, mask_generator, clip_model, clip_processor
|
| 156 |
|
| 157 |
-
if imagen is None:
|
| 158 |
-
return None, "
|
| 159 |
-
|
| 160 |
-
#
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
if sam2_model is None:
|
| 164 |
-
print("Inicializando modelos en GPU por primera vez...")
|
| 165 |
-
|
| 166 |
-
# Activar precisiones mixtas para acelerar
|
| 167 |
torch.autocast("cuda", dtype=torch.bfloat16).__enter__()
|
| 168 |
if torch.cuda.get_device_properties(0).major >= 8:
|
| 169 |
torch.backends.cuda.matmul.allow_tf32 = True
|
| 170 |
torch.backends.cudnn.allow_tf32 = True
|
| 171 |
|
|
|
|
|
|
|
| 172 |
sam2_model = build_sam2(SAM2_CONFIG, checkpoint_path, device=DEVICE)
|
| 173 |
-
|
| 174 |
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
|
| 179 |
-
# --- INICIO DEL PROCESAMIENTO ---
|
| 180 |
imagen = imagen.convert("RGB")
|
| 181 |
imagen_np = np.array(imagen)
|
| 182 |
-
|
| 183 |
-
masks = mask_generator.generate(imagen_np)
|
| 184 |
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 192 |
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
terms = [t.strip() for t in re.split(r"[,\n]+", texto) if t.strip()]
|
| 196 |
-
return None, f"No se encontró un objeto que coincida con: {', '.join(terms)}."
|
| 197 |
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
|
| 202 |
-
|
| 203 |
-
if missing_terms:
|
| 204 |
-
message += f" No se encontró: {', '.join(missing_terms)}."
|
| 205 |
-
return overlay, message
|
| 206 |
|
| 207 |
def crear_app():
|
| 208 |
-
with gr.Blocks(title="
|
| 209 |
-
gr.Markdown("#
|
| 210 |
gr.Markdown(
|
| 211 |
-
"
|
|
|
|
| 212 |
)
|
| 213 |
|
| 214 |
-
with gr.Row(
|
| 215 |
with gr.Column(scale=1):
|
| 216 |
-
imagen_entrada = gr.Image(type="pil", label="
|
| 217 |
-
texto_objeto = gr.Textbox(label="Buscar objeto", placeholder="Ej.
|
| 218 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 219 |
with gr.Column(scale=1):
|
| 220 |
-
imagen_salida = gr.Image(label="Resultado
|
| 221 |
estado = gr.Textbox(label="Estado", interactive=False)
|
| 222 |
|
| 223 |
boton.click(
|
| 224 |
-
fn=
|
| 225 |
-
inputs=[imagen_entrada, texto_objeto],
|
| 226 |
outputs=[imagen_salida, estado],
|
| 227 |
)
|
| 228 |
|
| 229 |
return demo
|
| 230 |
|
| 231 |
# --- INICIALIZACIÓN GLOBAL ---
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
checkpoint_path = download_checkpoint()
|
| 235 |
|
| 236 |
-
# Iniciar App (los modelos se cargarán al hacer clic en Segmentar)
|
| 237 |
demo = crear_app()
|
| 238 |
|
| 239 |
if __name__ == "__main__":
|
|
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
| 2 |
import gradio as gr
|
| 3 |
import numpy as np
|
| 4 |
import torch
|
| 5 |
+
from pathlib import Path
|
| 6 |
from huggingface_hub import hf_hub_download
|
| 7 |
from PIL import Image
|
|
|
|
| 8 |
import spaces # <-- Importante para Hugging Face ZeroGPU
|
| 9 |
|
| 10 |
+
# --- IMPORTACIONES DE MODELOS ---
|
| 11 |
+
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
|
| 12 |
from sam2.build_sam import build_sam2
|
| 13 |
+
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
| 14 |
|
| 15 |
+
# --- CONFIGURACIÓN DE MODELOS ---
|
| 16 |
+
# SAM 2.1
|
| 17 |
SAM2_REPO = "facebook/sam2.1-hiera-base-plus"
|
| 18 |
CHECKPOINT_FILENAME = "sam2.1_hiera_base_plus.pt"
|
| 19 |
SAM2_CONFIG = "configs/sam2.1/sam2.1_hiera_b+.yaml"
|
| 20 |
|
| 21 |
+
# GroundingDINO
|
| 22 |
+
GDINO_ID = "IDEA-Research/grounding-dino-base"
|
| 23 |
|
|
|
|
| 24 |
DEVICE = "cuda"
|
|
|
|
| 25 |
|
| 26 |
+
# Variables globales para Lazy Loading (ZeroGPU)
|
| 27 |
+
sam2_predictor = None
|
| 28 |
+
gdino_model = None
|
| 29 |
clip_model = None
|
| 30 |
clip_processor = None
|
| 31 |
|
| 32 |
COLOR_PALETTE = [
|
| 33 |
+
(0, 255, 255, 150), # Cian (queda muy bien para resaltar)
|
| 34 |
+
(255, 0, 255, 150), # Magenta
|
| 35 |
+
(255, 255, 0, 150), # Amarillo
|
| 36 |
+
(0, 255, 0, 150), # Verde
|
| 37 |
+
(255, 0, 0, 150), # Rojo
|
| 38 |
+
(0, 0, 255, 150), # Azul
|
| 39 |
]
|
| 40 |
|
| 41 |
+
def download_sam_checkpoint() -> str:
|
|
|
|
| 42 |
cache_dir = Path("./models")
|
| 43 |
cache_dir.mkdir(parents=True, exist_ok=True)
|
| 44 |
local_path = cache_dir / CHECKPOINT_FILENAME
|
|
|
|
| 45 |
if not local_path.exists():
|
| 46 |
+
print(f"Descargando {CHECKPOINT_FILENAME}...")
|
| 47 |
+
local_path = Path(hf_hub_download(repo_id=SAM2_REPO, filename=CHECKPOINT_FILENAME, cache_dir=str(cache_dir)))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
return str(local_path)
|
| 49 |
|
| 50 |
+
def create_mask_overlay(image: Image.Image, masks_np: np.ndarray) -> Image.Image:
|
| 51 |
+
"""Superpone las máscaras booleanas (N, H, W) sobre la imagen."""
|
| 52 |
+
overlay_image = image.convert("RGBA").copy()
|
| 53 |
+
|
| 54 |
+
for i, mask_bool in enumerate(masks_np):
|
|
|
|
|
|
|
|
|
|
| 55 |
color = COLOR_PALETTE[i % len(COLOR_PALETTE)]
|
| 56 |
+
mask_image = Image.fromarray((mask_bool * 255).astype(np.uint8), mode="L")
|
| 57 |
+
color_overlay = Image.new("RGBA", overlay_image.size, color)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
overlay_image.paste(color_overlay, (0, 0), mask_image)
|
| 59 |
|
| 60 |
return overlay_image
|
| 61 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
@spaces.GPU
|
| 63 |
@torch.no_grad()
|
| 64 |
+
def segmentar_con_dino_y_sam(imagen: Image.Image, texto: str, box_threshold: float):
|
| 65 |
+
global sam2_predictor, gdino_model, gdino_processor
|
|
|
|
| 66 |
|
| 67 |
+
if imagen is None or not texto.strip():
|
| 68 |
+
return None, "Sube una imagen y escribe qué quieres buscar."
|
| 69 |
+
|
| 70 |
+
# 1. LAZY LOADING: Inicializar modelos en la GPU la primera vez
|
| 71 |
+
if sam2_predictor is None:
|
| 72 |
+
print("Inicializando GroundingDINO y SAM 2.1 en GPU...")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
torch.autocast("cuda", dtype=torch.bfloat16).__enter__()
|
| 74 |
if torch.cuda.get_device_properties(0).major >= 8:
|
| 75 |
torch.backends.cuda.matmul.allow_tf32 = True
|
| 76 |
torch.backends.cudnn.allow_tf32 = True
|
| 77 |
|
| 78 |
+
# Cargar SAM 2.1 en modo Predictor (para cajas), no AutomaticMaskGenerator
|
| 79 |
+
checkpoint_path = download_sam_checkpoint()
|
| 80 |
sam2_model = build_sam2(SAM2_CONFIG, checkpoint_path, device=DEVICE)
|
| 81 |
+
sam2_predictor = SAM2ImagePredictor(sam2_model)
|
| 82 |
|
| 83 |
+
# Cargar GroundingDINO
|
| 84 |
+
gdino_processor = AutoProcessor.from_pretrained(GDINO_ID)
|
| 85 |
+
gdino_model = AutoModelForZeroShotObjectDetection.from_pretrained(GDINO_ID).to(DEVICE)
|
| 86 |
+
print("¡Modelos listos!")
|
| 87 |
+
|
| 88 |
+
# Asegurarnos de que el texto termine en punto (GroundingDINO funciona mejor así)
|
| 89 |
+
texto = texto.strip()
|
| 90 |
+
if not texto.endswith("."):
|
| 91 |
+
texto += "."
|
| 92 |
|
|
|
|
| 93 |
imagen = imagen.convert("RGB")
|
| 94 |
imagen_np = np.array(imagen)
|
|
|
|
|
|
|
| 95 |
|
| 96 |
+
# 2. GROUNDING DINO: Encontrar las cajas delimitadoras
|
| 97 |
+
inputs = gdino_processor(images=imagen, text=texto, return_tensors="pt").to(DEVICE)
|
| 98 |
+
outputs = gdino_model(**inputs)
|
| 99 |
+
|
| 100 |
+
# Extraer las cajas con un umbral de confianza
|
| 101 |
+
results = gdino_processor.post_process_grounded_object_detection(
|
| 102 |
+
outputs,
|
| 103 |
+
inputs.input_ids,
|
| 104 |
+
box_threshold=box_threshold,
|
| 105 |
+
text_threshold=0.25,
|
| 106 |
+
target_sizes=[imagen.size[::-1]] # (alto, ancho)
|
| 107 |
+
)[0]
|
| 108 |
+
|
| 109 |
+
cajas = results["boxes"] # Tensor con coordenadas [x1, y1, x2, y2]
|
| 110 |
+
etiquetas = results["labels"]
|
| 111 |
+
scores = results["scores"]
|
| 112 |
+
|
| 113 |
+
if len(cajas) == 0:
|
| 114 |
+
return imagen, f"No se encontró nada para '{texto}' con el umbral actual ({box_threshold}). Intenta bajarlo."
|
| 115 |
+
|
| 116 |
+
# 3. SAM 2.1: Segmentar dentro de las cajas encontradas
|
| 117 |
+
sam2_predictor.set_image(imagen_np)
|
| 118 |
+
|
| 119 |
+
# SAM 2.1 requiere que las cajas sean un array numpy
|
| 120 |
+
input_boxes = cajas.cpu().numpy()
|
| 121 |
+
|
| 122 |
+
masks, _, _ = sam2_predictor.predict(
|
| 123 |
+
point_coords=None,
|
| 124 |
+
point_labels=None,
|
| 125 |
+
box=input_boxes,
|
| 126 |
+
multimask_output=False, # Queremos 1 máscara final por caja
|
| 127 |
+
)
|
| 128 |
+
|
| 129 |
+
# Las máscaras de SAM tienen forma (N, 1, H, W). Las aplanamos a (N, H, W)
|
| 130 |
+
masks = masks.squeeze(1)
|
| 131 |
|
| 132 |
+
# 4. SUPERPONER MÁSCARAS
|
| 133 |
+
resultado_img = create_mask_overlay(imagen, masks)
|
|
|
|
|
|
|
| 134 |
|
| 135 |
+
# Preparar el mensaje de estado
|
| 136 |
+
objetos_encontrados = [f"{label} ({score:.2f})" for label, score in zip(etiquetas, scores)]
|
| 137 |
+
mensaje = f"Encontrados {len(cajas)} objeto(s): {', '.join(objetos_encontrados)}"
|
| 138 |
|
| 139 |
+
return resultado_img, mensaje
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
def crear_app():
|
| 142 |
+
with gr.Blocks(title="GroundingDINO + SAM 2.1") as demo:
|
| 143 |
+
gr.Markdown("# 🦖 GroundingDINO + 🎯 SAM 2.1 (Base Plus)")
|
| 144 |
gr.Markdown(
|
| 145 |
+
"Segmentación de alta precisión basada en texto. Escribe lo que buscas (ej. `bed`, `lamp`, `pillow`).\n\n"
|
| 146 |
+
"*Nota: La primera imagen tardará unos segundos mientras se inicializa la GPU.*"
|
| 147 |
)
|
| 148 |
|
| 149 |
+
with gr.Row():
|
| 150 |
with gr.Column(scale=1):
|
| 151 |
+
imagen_entrada = gr.Image(type="pil", label="Sube tu foto")
|
| 152 |
+
texto_objeto = gr.Textbox(label="Buscar objeto (en inglés funciona mejor)", placeholder="Ej. bed, pillow, carpet")
|
| 153 |
+
|
| 154 |
+
# Deslizador para ajustar la sensibilidad de GroundingDINO
|
| 155 |
+
umbral = gr.Slider(minimum=0.1, maximum=0.9, value=0.3, step=0.05, label="Umbral de detección (Box Threshold)", info="Bájalo si no detecta el objeto, súbelo si detecta cosas incorrectas.")
|
| 156 |
+
|
| 157 |
+
boton = gr.Button("Segmentar", variant="primary")
|
| 158 |
+
|
| 159 |
with gr.Column(scale=1):
|
| 160 |
+
imagen_salida = gr.Image(label="Resultado Segmentado")
|
| 161 |
estado = gr.Textbox(label="Estado", interactive=False)
|
| 162 |
|
| 163 |
boton.click(
|
| 164 |
+
fn=segmentar_con_dino_y_sam,
|
| 165 |
+
inputs=[imagen_entrada, texto_objeto, umbral],
|
| 166 |
outputs=[imagen_salida, estado],
|
| 167 |
)
|
| 168 |
|
| 169 |
return demo
|
| 170 |
|
| 171 |
# --- INICIALIZACIÓN GLOBAL ---
|
| 172 |
+
print("Descargando peso de SAM 2.1 al iniciar Space...")
|
| 173 |
+
download_sam_checkpoint()
|
|
|
|
| 174 |
|
|
|
|
| 175 |
demo = crear_app()
|
| 176 |
|
| 177 |
if __name__ == "__main__":
|