Spaces:
Running
Running
Commit
·
4e07023
1
Parent(s):
0970982
Deploy: Versión inicial con modelos v4
Browse files- main.py +118 -0
- models/beto_complex_v4.pth +3 -0
- models/beto_simple_v4.pth +3 -0
- packages.txt +2 -0
- requeriments.txt +13 -0
- requirements.txt +0 -3
- src/__init__.py +0 -0
- src/__pycache__/inference.cpython-313.pyc +0 -0
- src/__pycache__/nlp_utils.cpython-313.pyc +0 -0
- src/__pycache__/utils.cpython-313.pyc +0 -0
- src/directory.py +62 -0
- src/inference.py +137 -0
- src/nlp_utils.py +46 -0
- src/split-data.py +75 -0
- src/streamlit_app.py +0 -40
- src/utils.py +51 -0
main.py
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import streamlit as st
|
| 2 |
+
import pandas as pd
|
| 3 |
+
from PIL import Image
|
| 4 |
+
import sys
|
| 5 |
+
import os
|
| 6 |
+
|
| 7 |
+
# Configuración de rutas
|
| 8 |
+
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '/src')))
|
| 9 |
+
from inference import MemePredictor
|
| 10 |
+
|
| 11 |
+
# --- 1. CONFIGURACIÓN DE PÁGINA (Minimalista) ---
|
| 12 |
+
st.set_page_config(
|
| 13 |
+
page_title="DIME-MEX",
|
| 14 |
+
layout="centered",
|
| 15 |
+
initial_sidebar_state="collapsed"
|
| 16 |
+
)
|
| 17 |
+
|
| 18 |
+
# --- 2. ESTILOS CSS PERSONALIZADOS (Para móvil) ---
|
| 19 |
+
st.markdown("""
|
| 20 |
+
<style>
|
| 21 |
+
/* Ocultar menú hamburguesa y footer para limpieza visual */
|
| 22 |
+
#MainMenu {visibility: hidden;}
|
| 23 |
+
footer {visibility: hidden;}
|
| 24 |
+
|
| 25 |
+
/* Ajustar padding para móviles */
|
| 26 |
+
.block-container {
|
| 27 |
+
padding-top: 2rem;
|
| 28 |
+
padding-bottom: 2rem;
|
| 29 |
+
}
|
| 30 |
+
|
| 31 |
+
/* Botón primario minimalista (Blanco y Negro o color acento del tema) */
|
| 32 |
+
div.stButton > button:first-child {
|
| 33 |
+
width: 100%;
|
| 34 |
+
border-radius: 5px;
|
| 35 |
+
height: 3em;
|
| 36 |
+
font-weight: bold;
|
| 37 |
+
}
|
| 38 |
+
</style>
|
| 39 |
+
""", unsafe_allow_html=True)
|
| 40 |
+
|
| 41 |
+
# --- 3. CARGA DEL MOTOR ---
|
| 42 |
+
@st.cache_resource
|
| 43 |
+
def get_engine():
|
| 44 |
+
return MemePredictor()
|
| 45 |
+
|
| 46 |
+
try:
|
| 47 |
+
predictor = get_engine()
|
| 48 |
+
except Exception as e:
|
| 49 |
+
st.error(f"Error iniciando el sistema: {e}")
|
| 50 |
+
st.stop()
|
| 51 |
+
|
| 52 |
+
# --- 4. INTERFAZ DE USUARIO ---
|
| 53 |
+
|
| 54 |
+
# Título limpio
|
| 55 |
+
st.markdown("## Detector de Contenido")
|
| 56 |
+
st.markdown("Clasificación de memes utilizando visión y lenguaje natural.")
|
| 57 |
+
|
| 58 |
+
# ACORDEÓN DE CONFIGURACIÓN
|
| 59 |
+
with st.expander("Configuración del Modelo"):
|
| 60 |
+
task_mode = st.radio(
|
| 61 |
+
"Nivel de detalle:",
|
| 62 |
+
options=["simple", "complex"],
|
| 63 |
+
format_func=lambda x: "Básico (3 Categorías)" if x == "simple" else "Detallado (6 Categorías)"
|
| 64 |
+
)
|
| 65 |
+
|
| 66 |
+
# ÁREA DE DRAG AND DROP
|
| 67 |
+
uploaded_file = st.file_uploader("Cargar imagen", type=["jpg", "png", "jpeg"], help="Arrastra tu archivo aquí")
|
| 68 |
+
|
| 69 |
+
if uploaded_file is not None:
|
| 70 |
+
# Mostrar imagen centrada y ajustada al ancho del móvil
|
| 71 |
+
image = Image.open(uploaded_file)
|
| 72 |
+
st.image(image, use_column_width=True)
|
| 73 |
+
|
| 74 |
+
# Botón de acción (ocupa todo el ancho por el CSS)
|
| 75 |
+
if st.button("ANALIZAR IMAGEN", type="primary"):
|
| 76 |
+
|
| 77 |
+
with st.spinner("Procesando..."):
|
| 78 |
+
try:
|
| 79 |
+
# Inferencia
|
| 80 |
+
result = predictor.predict(uploaded_file, task=task_mode)
|
| 81 |
+
|
| 82 |
+
if "error" in result:
|
| 83 |
+
st.error(result["error"])
|
| 84 |
+
else:
|
| 85 |
+
st.divider() # Línea separadora sutil
|
| 86 |
+
|
| 87 |
+
# --- RESULTADOS MINIMALISTAS (st.metric) ---
|
| 88 |
+
# Usamos columnas para que se vea bien en celular (uno al lado del otro o apilados)
|
| 89 |
+
col1, col2 = st.columns(2)
|
| 90 |
+
|
| 91 |
+
with col1:
|
| 92 |
+
st.metric(label="Clasificación", value=result['label'])
|
| 93 |
+
|
| 94 |
+
with col2:
|
| 95 |
+
# Formato de porcentaje limpio
|
| 96 |
+
st.metric(label="Confianza", value=f"{result['confidence']:.1%}")
|
| 97 |
+
|
| 98 |
+
# Barra de progreso simple
|
| 99 |
+
st.progress(result['confidence'])
|
| 100 |
+
|
| 101 |
+
# --- DETALLES TÉCNICOS (Ocultos por defecto) ---
|
| 102 |
+
with st.expander("Ver texto extraído"):
|
| 103 |
+
st.caption("Texto detectado por OCR:")
|
| 104 |
+
st.text(result["ocr_text"])
|
| 105 |
+
st.caption("Texto procesado para el modelo:")
|
| 106 |
+
st.code(result["clean_text"], language="text")
|
| 107 |
+
|
| 108 |
+
# --- GRÁFICA LIMPIA ---
|
| 109 |
+
st.caption("Distribución de probabilidades")
|
| 110 |
+
chart_data = pd.DataFrame({
|
| 111 |
+
"Categoría": result['all_labels'],
|
| 112 |
+
"Probabilidad": result['probabilities']
|
| 113 |
+
})
|
| 114 |
+
# Gráfica de barras horizontal es mejor para leer etiquetas largas en móvil
|
| 115 |
+
st.bar_chart(chart_data.set_index("Categoría"), color="#333333")
|
| 116 |
+
|
| 117 |
+
except Exception as e:
|
| 118 |
+
st.error(f"Error interno: {e}")
|
models/beto_complex_v4.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6041335c6d5f8317d944157ecc99d801d5d069c429923856670275af7fe62073
|
| 3 |
+
size 439506051
|
models/beto_simple_v4.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dd13bd86b8e9ea6f5f1dcbab9abaa75cf23552c370a6badb509322d0043f19a0
|
| 3 |
+
size 439496628
|
packages.txt
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
libgl1-mesa-glx
|
| 2 |
+
libglib2.0-0
|
requeriments.txt
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
pandas
|
| 2 |
+
numpy
|
| 3 |
+
scikit-learn
|
| 4 |
+
torch
|
| 5 |
+
huggingface-hub<1.0
|
| 6 |
+
transformers
|
| 7 |
+
nltk
|
| 8 |
+
matplotlib
|
| 9 |
+
seaborn
|
| 10 |
+
tqdm
|
| 11 |
+
streamlit
|
| 12 |
+
easyocr
|
| 13 |
+
opencv-python-headless
|
requirements.txt
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
altair
|
| 2 |
-
pandas
|
| 3 |
-
streamlit
|
|
|
|
|
|
|
|
|
|
|
|
src/__init__.py
ADDED
|
File without changes
|
src/__pycache__/inference.cpython-313.pyc
ADDED
|
Binary file (6.2 kB). View file
|
|
|
src/__pycache__/nlp_utils.cpython-313.pyc
ADDED
|
Binary file (3.27 kB). View file
|
|
|
src/__pycache__/utils.cpython-313.pyc
ADDED
|
Binary file (2.64 kB). View file
|
|
|
src/directory.py
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import pandas as pd
|
| 2 |
+
import json
|
| 3 |
+
import os
|
| 4 |
+
|
| 5 |
+
# Configuracion de rutas dinamicas
|
| 6 |
+
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
|
| 7 |
+
PROJECT_ROOT = os.path.abspath(os.path.join(SCRIPT_DIR, "../"))
|
| 8 |
+
|
| 9 |
+
# Directorios de entrada y salida
|
| 10 |
+
INPUT_DIR = os.path.join(PROJECT_ROOT, "data", "train")
|
| 11 |
+
OUTPUT_DIR = os.path.join(PROJECT_ROOT, "data", "processed", "datasets")
|
| 12 |
+
|
| 13 |
+
TASKS = {
|
| 14 |
+
"simple": {
|
| 15 |
+
"file": "label_simple.csv",
|
| 16 |
+
"cols": ["none", "inappropriate", "hate_speech"]
|
| 17 |
+
},
|
| 18 |
+
"complex": {
|
| 19 |
+
"file": "label_complex.csv",
|
| 20 |
+
"cols": ["none", "inappropriate", "sexism", "racism", "classicism", "other"]
|
| 21 |
+
}
|
| 22 |
+
}
|
| 23 |
+
|
| 24 |
+
def create_dataset_csv(task_name):
|
| 25 |
+
# Validar existencia del directorio de entrada
|
| 26 |
+
if not os.path.exists(INPUT_DIR):
|
| 27 |
+
print(f"Error: No se encuentra el directorio {INPUT_DIR}")
|
| 28 |
+
return
|
| 29 |
+
|
| 30 |
+
# Cargar archivo JSON con metadatos
|
| 31 |
+
json_path = os.path.join(INPUT_DIR, "train_data.json")
|
| 32 |
+
with open(json_path, 'r', encoding='utf-8') as f:
|
| 33 |
+
data_json = json.load(f)
|
| 34 |
+
|
| 35 |
+
df = pd.DataFrame(data_json)
|
| 36 |
+
|
| 37 |
+
# Cargar archivo CSV con etiquetas
|
| 38 |
+
csv_conf = TASKS[task_name]
|
| 39 |
+
labels_path = os.path.join(INPUT_DIR, csv_conf["file"])
|
| 40 |
+
df_labels = pd.read_csv(labels_path, header=None)
|
| 41 |
+
|
| 42 |
+
# Convertir one-hot encoding a etiquetas numericas
|
| 43 |
+
df['label'] = df_labels.values.argmax(axis=1)
|
| 44 |
+
|
| 45 |
+
# Generar rutas relativas de las imagenes
|
| 46 |
+
images_prefix = os.path.join("data", "train", "images")
|
| 47 |
+
df['path'] = df['MEME-ID'].apply(lambda x: os.path.join(images_prefix, x))
|
| 48 |
+
|
| 49 |
+
# Seleccionar columnas finales
|
| 50 |
+
final_df = df[['path', 'text', 'description', 'label']]
|
| 51 |
+
|
| 52 |
+
# Crear directorio de salida si no existe
|
| 53 |
+
os.makedirs(OUTPUT_DIR, exist_ok=True)
|
| 54 |
+
|
| 55 |
+
# Guardar archivo CSV
|
| 56 |
+
output_file = os.path.join(OUTPUT_DIR, f"dataset-{task_name}.csv")
|
| 57 |
+
final_df.to_csv(output_file, index=False)
|
| 58 |
+
print(f"Archivo generado: {output_file}")
|
| 59 |
+
|
| 60 |
+
if __name__ == "__main__":
|
| 61 |
+
create_dataset_csv("simple")
|
| 62 |
+
create_dataset_csv("complex")
|
src/inference.py
ADDED
|
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import easyocr
|
| 3 |
+
import os
|
| 4 |
+
import sys
|
| 5 |
+
from transformers import AutoTokenizer
|
| 6 |
+
|
| 7 |
+
# Configurar path para importar módulos locales (nlp_utils y utils)
|
| 8 |
+
CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))
|
| 9 |
+
sys.path.append(CURRENT_DIR)
|
| 10 |
+
|
| 11 |
+
from nlp_utils import BetoClassifier
|
| 12 |
+
from utils import clean_text, preprocess_image_for_ocr
|
| 13 |
+
|
| 14 |
+
# ==========================================
|
| 15 |
+
# ⚙️ CONFIGURACIÓN DEL MODELO
|
| 16 |
+
# ==========================================
|
| 17 |
+
MODEL_VERSION = "v4" # <--- CAMBIA ESTO si entrenas nuevas versiones
|
| 18 |
+
MODEL_NAME = "dccuchile/bert-base-spanish-wwm-cased"
|
| 19 |
+
MAX_LEN = 128
|
| 20 |
+
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 21 |
+
|
| 22 |
+
# Construcción de la ruta: .../DIMEMEX/models/v4/
|
| 23 |
+
# Subimos dos niveles desde src/inference.py para llegar a la raiz, luego models, luego v4
|
| 24 |
+
PROJECT_ROOT = os.path.abspath(os.path.join(CURRENT_DIR, "../"))
|
| 25 |
+
MODEL_DIR = os.path.join(PROJECT_ROOT, "models")
|
| 26 |
+
|
| 27 |
+
class MemePredictor:
|
| 28 |
+
def __init__(self):
|
| 29 |
+
print(f"🔧 Inicializando motor en: {DEVICE}")
|
| 30 |
+
print(f"📂 Buscando modelos versión {MODEL_VERSION} en: {MODEL_DIR}")
|
| 31 |
+
|
| 32 |
+
# Cargar OCR (Singleton)
|
| 33 |
+
self.reader = easyocr.Reader(['es', 'en'], gpu=(DEVICE.type == 'cuda'))
|
| 34 |
+
self.tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 35 |
+
|
| 36 |
+
# Cache para no recargar modelos .pth
|
| 37 |
+
self.loaded_models = {}
|
| 38 |
+
|
| 39 |
+
# Mapas de etiquetas
|
| 40 |
+
self.labels_map = {
|
| 41 |
+
"simple": ["None", "Inappropriate", "Hate"],
|
| 42 |
+
"complex": ["None", "Inapp", "Sexism", "Racism", "Classicism", "Other"]
|
| 43 |
+
}
|
| 44 |
+
|
| 45 |
+
def _get_model_instance(self, task):
|
| 46 |
+
# Si ya está en RAM, devolverlo
|
| 47 |
+
if task in self.loaded_models:
|
| 48 |
+
return self.loaded_models[task]
|
| 49 |
+
|
| 50 |
+
# Construir nombre del archivo: beto_simple_v4.pth
|
| 51 |
+
filename = f"beto_{task}_{MODEL_VERSION}.pth"
|
| 52 |
+
path = os.path.join(MODEL_DIR, filename)
|
| 53 |
+
|
| 54 |
+
if not os.path.exists(path):
|
| 55 |
+
raise FileNotFoundError(f"No se encontró el modelo para la tarea '{task}' en {path}")
|
| 56 |
+
|
| 57 |
+
print(f"📥 Cargando modelo desde: {path}")
|
| 58 |
+
|
| 59 |
+
n_classes = len(self.labels_map[task])
|
| 60 |
+
|
| 61 |
+
# Instanciar arquitectura
|
| 62 |
+
model = BetoClassifier(n_classes, MODEL_NAME)
|
| 63 |
+
|
| 64 |
+
# Cargar pesos
|
| 65 |
+
# map_location es vital para evitar errores si entrenaste en GPU y corres en CPU
|
| 66 |
+
try:
|
| 67 |
+
model.load_state_dict(torch.load(path, map_location=DEVICE))
|
| 68 |
+
except Exception as e:
|
| 69 |
+
raise RuntimeError(f"Error al leer el archivo .pth: {e}")
|
| 70 |
+
|
| 71 |
+
model.to(DEVICE)
|
| 72 |
+
model.eval()
|
| 73 |
+
|
| 74 |
+
self.loaded_models[task] = model
|
| 75 |
+
return model
|
| 76 |
+
|
| 77 |
+
def predict(self, image_file, task="simple"):
|
| 78 |
+
# 1. Resetear puntero del archivo
|
| 79 |
+
image_file.seek(0)
|
| 80 |
+
|
| 81 |
+
# 2. Pre-procesamiento de Imagen (Visión)
|
| 82 |
+
proc_img, raw_img = preprocess_image_for_ocr(image_file)
|
| 83 |
+
|
| 84 |
+
if proc_img is None:
|
| 85 |
+
return {"error": "Error procesando la imagen (archivo corrupto o formato inválido)"}
|
| 86 |
+
|
| 87 |
+
# 3. OCR con Fallback
|
| 88 |
+
try:
|
| 89 |
+
# Intento 1: Imagen procesada
|
| 90 |
+
ocr_result = self.reader.readtext(proc_img, detail=0, paragraph=True)
|
| 91 |
+
raw_text = " ".join(ocr_result)
|
| 92 |
+
|
| 93 |
+
# Si leyó muy poco (<3 chars), intentar con la original
|
| 94 |
+
if len(raw_text) < 3:
|
| 95 |
+
ocr_result = self.reader.readtext(raw_img, detail=0, paragraph=True)
|
| 96 |
+
raw_text = " ".join(ocr_result)
|
| 97 |
+
except Exception as e:
|
| 98 |
+
return {"error": f"Fallo en OCR: {e}"}
|
| 99 |
+
|
| 100 |
+
# 4. Limpieza de Texto (NLP)
|
| 101 |
+
text_ready = clean_text(raw_text)
|
| 102 |
+
|
| 103 |
+
# 5. Tokenización
|
| 104 |
+
encoding = self.tokenizer.encode_plus(
|
| 105 |
+
text_ready,
|
| 106 |
+
add_special_tokens=True,
|
| 107 |
+
max_length=MAX_LEN,
|
| 108 |
+
padding='max_length',
|
| 109 |
+
truncation=True,
|
| 110 |
+
return_attention_mask=True,
|
| 111 |
+
return_tensors='pt',
|
| 112 |
+
)
|
| 113 |
+
|
| 114 |
+
input_ids = encoding['input_ids'].to(DEVICE)
|
| 115 |
+
attention_mask = encoding['attention_mask'].to(DEVICE)
|
| 116 |
+
|
| 117 |
+
# 6. Inferencia del Modelo
|
| 118 |
+
try:
|
| 119 |
+
model = self._get_model_instance(task)
|
| 120 |
+
|
| 121 |
+
with torch.no_grad():
|
| 122 |
+
outputs = model(input_ids, attention_mask)
|
| 123 |
+
probs = torch.nn.functional.softmax(outputs, dim=1)
|
| 124 |
+
conf, idx = torch.max(probs, dim=1)
|
| 125 |
+
|
| 126 |
+
label_str = self.labels_map[task][idx.item()]
|
| 127 |
+
|
| 128 |
+
return {
|
| 129 |
+
"ocr_text": raw_text,
|
| 130 |
+
"clean_text": text_ready,
|
| 131 |
+
"label": label_str,
|
| 132 |
+
"confidence": conf.item(),
|
| 133 |
+
"probabilities": probs.cpu().numpy()[0],
|
| 134 |
+
"all_labels": self.labels_map[task]
|
| 135 |
+
}
|
| 136 |
+
except Exception as e:
|
| 137 |
+
return {"error": f"Error en inferencia del modelo: {e}"}
|
src/nlp_utils.py
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# src/nlp_utils.py
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
from torch.utils.data import Dataset
|
| 4 |
+
import torch
|
| 5 |
+
from transformers import AutoModel
|
| 6 |
+
|
| 7 |
+
class BetoClassifier(nn.Module):
|
| 8 |
+
def __init__(self, n_classes, model_name="dccuchile/bert-base-spanish-wwm-cased"):
|
| 9 |
+
super(BetoClassifier, self).__init__()
|
| 10 |
+
self.bert = AutoModel.from_pretrained(model_name)
|
| 11 |
+
self.drop = nn.Dropout(p=0.3)
|
| 12 |
+
self.out = nn.Linear(self.bert.config.hidden_size, n_classes)
|
| 13 |
+
|
| 14 |
+
def forward(self, input_ids, attention_mask):
|
| 15 |
+
output = self.bert(input_ids=input_ids, attention_mask=attention_mask)
|
| 16 |
+
return self.out(self.drop(output.pooler_output))
|
| 17 |
+
|
| 18 |
+
class MemeDataset(Dataset):
|
| 19 |
+
def __init__(self, df, tokenizer, max_len, label_col):
|
| 20 |
+
self.df = df.reset_index(drop=True)
|
| 21 |
+
self.tokenizer = tokenizer
|
| 22 |
+
self.max_len = max_len
|
| 23 |
+
self.label_col = label_col # Columna dinamica segun la tarea
|
| 24 |
+
|
| 25 |
+
def __len__(self):
|
| 26 |
+
return len(self.df)
|
| 27 |
+
|
| 28 |
+
def __getitem__(self, index):
|
| 29 |
+
text = str(self.df.loc[index, 'text_clean'])
|
| 30 |
+
label = int(self.df.loc[index, self.label_col])
|
| 31 |
+
|
| 32 |
+
encoding = self.tokenizer.encode_plus(
|
| 33 |
+
text,
|
| 34 |
+
add_special_tokens=True,
|
| 35 |
+
max_length=self.max_len,
|
| 36 |
+
padding='max_length',
|
| 37 |
+
truncation=True,
|
| 38 |
+
return_attention_mask=True,
|
| 39 |
+
return_tensors='pt',
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
return {
|
| 43 |
+
'input_ids': encoding['input_ids'].flatten(),
|
| 44 |
+
'attention_mask': encoding['attention_mask'].flatten(),
|
| 45 |
+
'label': torch.tensor(label, dtype=torch.long)
|
| 46 |
+
}
|
src/split-data.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import pandas as pd
|
| 2 |
+
from sklearn.model_selection import train_test_split
|
| 3 |
+
import os
|
| 4 |
+
|
| 5 |
+
# Configuracion de rutas dinamicas
|
| 6 |
+
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
|
| 7 |
+
PROJECT_ROOT = os.path.abspath(os.path.join(SCRIPT_DIR, "../"))
|
| 8 |
+
|
| 9 |
+
# Directorios de entrada y salida
|
| 10 |
+
DATA_PROCESSED_DIR = os.path.join(PROJECT_ROOT, "data", "processed","nlp")
|
| 11 |
+
TEXT_FILE_PATH = os.path.join(DATA_PROCESSED_DIR, "cleaned-ocr.csv")
|
| 12 |
+
SIMPLE_FILE_PATH = os.path.join(DATA_PROCESSED_DIR, "datasets", "dataset-simple.csv")
|
| 13 |
+
COMPLEX_FILE_PATH = os.path.join(DATA_PROCESSED_DIR, "datasets", "dataset-complex.csv")
|
| 14 |
+
OUTPUT_DIR = os.path.join(DATA_PROCESSED_DIR, "splits")
|
| 15 |
+
|
| 16 |
+
def create_master_splits():
|
| 17 |
+
# Validar existencia de archivos de entrada
|
| 18 |
+
if not all(os.path.exists(p) for p in [TEXT_FILE_PATH, SIMPLE_FILE_PATH, COMPLEX_FILE_PATH]):
|
| 19 |
+
print("Error: Faltan archivos de entrada en data/processed")
|
| 20 |
+
return
|
| 21 |
+
|
| 22 |
+
# Cargar archivos CSV
|
| 23 |
+
df_text = pd.read_csv(TEXT_FILE_PATH)
|
| 24 |
+
df_simple = pd.read_csv(SIMPLE_FILE_PATH)
|
| 25 |
+
df_complex = pd.read_csv(COMPLEX_FILE_PATH)
|
| 26 |
+
|
| 27 |
+
# Generar columna id basada en el nombre del archivo si no existe
|
| 28 |
+
if 'id' not in df_simple.columns:
|
| 29 |
+
df_simple['id'] = df_simple['path'].apply(os.path.basename)
|
| 30 |
+
if 'id' not in df_complex.columns:
|
| 31 |
+
df_complex['id'] = df_complex['path'].apply(os.path.basename)
|
| 32 |
+
|
| 33 |
+
# Renombrar columnas de etiquetas para evitar conflictos
|
| 34 |
+
df_simple = df_simple.rename(columns={'label': 'label-simple'})
|
| 35 |
+
df_complex = df_complex.rename(columns={'label': 'label-complex'})
|
| 36 |
+
|
| 37 |
+
# Unir etiquetas simples y complejas usando el id
|
| 38 |
+
df_labels_merged = pd.merge(df_simple, df_complex[['id', 'label-complex']], on='id', how='inner')
|
| 39 |
+
|
| 40 |
+
# Unir con el texto limpio para crear el dataset maestro
|
| 41 |
+
df_master = pd.merge(df_labels_merged, df_text[['id', 'text_clean']], on='id', how='inner')
|
| 42 |
+
|
| 43 |
+
# Crear directorio de salida si no existe
|
| 44 |
+
os.makedirs(OUTPUT_DIR, exist_ok=True)
|
| 45 |
+
|
| 46 |
+
# Guardar el dataset completo para entrenamiento final
|
| 47 |
+
complete_data_path = os.path.join(OUTPUT_DIR, "complete-data.csv")
|
| 48 |
+
df_master.to_csv(complete_data_path, index=False)
|
| 49 |
+
print(f"Dataset completo guardado en: {complete_data_path}")
|
| 50 |
+
|
| 51 |
+
# Separar conjunto de prueba manteniendo balance de clases complejas
|
| 52 |
+
df_temp, df_test = train_test_split(
|
| 53 |
+
df_master,
|
| 54 |
+
test_size=0.15,
|
| 55 |
+
random_state=42,
|
| 56 |
+
stratify=df_master['label-complex']
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
# Separar entrenamiento y validacion del conjunto restante
|
| 60 |
+
df_train, df_val = train_test_split(
|
| 61 |
+
df_temp,
|
| 62 |
+
test_size=0.176,
|
| 63 |
+
random_state=42,
|
| 64 |
+
stratify=df_temp['label-complex']
|
| 65 |
+
)
|
| 66 |
+
|
| 67 |
+
# Guardar los archivos de particion
|
| 68 |
+
df_train.to_csv(os.path.join(OUTPUT_DIR, "train.csv"), index=False)
|
| 69 |
+
df_val.to_csv(os.path.join(OUTPUT_DIR, "val.csv"), index=False)
|
| 70 |
+
df_test.to_csv(os.path.join(OUTPUT_DIR, "test.csv"), index=False)
|
| 71 |
+
|
| 72 |
+
print(f"Splits guardados en: {OUTPUT_DIR}")
|
| 73 |
+
|
| 74 |
+
if __name__ == "__main__":
|
| 75 |
+
create_master_splits()
|
src/streamlit_app.py
DELETED
|
@@ -1,40 +0,0 @@
|
|
| 1 |
-
import altair as alt
|
| 2 |
-
import numpy as np
|
| 3 |
-
import pandas as pd
|
| 4 |
-
import streamlit as st
|
| 5 |
-
|
| 6 |
-
"""
|
| 7 |
-
# Welcome to Streamlit!
|
| 8 |
-
|
| 9 |
-
Edit `/streamlit_app.py` to customize this app to your heart's desire :heart:.
|
| 10 |
-
If you have any questions, checkout our [documentation](https://docs.streamlit.io) and [community
|
| 11 |
-
forums](https://discuss.streamlit.io).
|
| 12 |
-
|
| 13 |
-
In the meantime, below is an example of what you can do with just a few lines of code:
|
| 14 |
-
"""
|
| 15 |
-
|
| 16 |
-
num_points = st.slider("Number of points in spiral", 1, 10000, 1100)
|
| 17 |
-
num_turns = st.slider("Number of turns in spiral", 1, 300, 31)
|
| 18 |
-
|
| 19 |
-
indices = np.linspace(0, 1, num_points)
|
| 20 |
-
theta = 2 * np.pi * num_turns * indices
|
| 21 |
-
radius = indices
|
| 22 |
-
|
| 23 |
-
x = radius * np.cos(theta)
|
| 24 |
-
y = radius * np.sin(theta)
|
| 25 |
-
|
| 26 |
-
df = pd.DataFrame({
|
| 27 |
-
"x": x,
|
| 28 |
-
"y": y,
|
| 29 |
-
"idx": indices,
|
| 30 |
-
"rand": np.random.randn(num_points),
|
| 31 |
-
})
|
| 32 |
-
|
| 33 |
-
st.altair_chart(alt.Chart(df, height=700, width=700)
|
| 34 |
-
.mark_point(filled=True)
|
| 35 |
-
.encode(
|
| 36 |
-
x=alt.X("x", axis=None),
|
| 37 |
-
y=alt.Y("y", axis=None),
|
| 38 |
-
color=alt.Color("idx", legend=None, scale=alt.Scale()),
|
| 39 |
-
size=alt.Size("rand", legend=None, scale=alt.Scale(range=[1, 150])),
|
| 40 |
-
))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/utils.py
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import re
|
| 2 |
+
import cv2
|
| 3 |
+
import numpy as np
|
| 4 |
+
import pandas as pd
|
| 5 |
+
|
| 6 |
+
def clean_text(text):
|
| 7 |
+
"""Limpieza estándar para BERT."""
|
| 8 |
+
if not text or pd.isna(text): return "sin texto"
|
| 9 |
+
text = str(text).lower()
|
| 10 |
+
# Eliminar URLs y usuarios
|
| 11 |
+
text = re.sub(r'http\S+|www\.\S+', '', text)
|
| 12 |
+
text = re.sub(r'@\w+', '', text)
|
| 13 |
+
# Normalizar risas
|
| 14 |
+
text = re.sub(r'(ja|je|ha|he|lo){2,}', 'jaja', text)
|
| 15 |
+
# Eliminar basura de OCR
|
| 16 |
+
text = re.sub(r'[|_~*^>\[\]]', ' ', text)
|
| 17 |
+
# Espacios y saltos
|
| 18 |
+
text = text.replace('\n', ' ').replace('\r', ' ')
|
| 19 |
+
text = re.sub(r'\s+', ' ', text).strip()
|
| 20 |
+
return text if text else "sin texto"
|
| 21 |
+
|
| 22 |
+
def preprocess_image_for_ocr(file_bytes):
|
| 23 |
+
"""
|
| 24 |
+
Recibe bytes (desde Streamlit) y aplica filtros de OpenCV.
|
| 25 |
+
Retorna: (imagen_binarizada, imagen_original_cv2)
|
| 26 |
+
"""
|
| 27 |
+
# Convertir bytes a array numpy para OpenCV
|
| 28 |
+
file_bytes = np.asarray(bytearray(file_bytes.read()), dtype=np.uint8)
|
| 29 |
+
img = cv2.imdecode(file_bytes, 1) # 1 = Color BGR
|
| 30 |
+
|
| 31 |
+
if img is None: return None, None
|
| 32 |
+
|
| 33 |
+
# Pipeline de Mejora (Igual al benchmark)
|
| 34 |
+
try:
|
| 35 |
+
# 1. Upscaling (2x)
|
| 36 |
+
img_resized = cv2.resize(img, None, fx=2, fy=2, interpolation=cv2.INTER_CUBIC)
|
| 37 |
+
|
| 38 |
+
# 2. Escala de Grises
|
| 39 |
+
gray = cv2.cvtColor(img_resized, cv2.COLOR_BGR2GRAY)
|
| 40 |
+
|
| 41 |
+
# 3. Denoising
|
| 42 |
+
denoised = cv2.fastNlMeansDenoising(gray, None, h=10, templateWindowSize=7, searchWindowSize=21)
|
| 43 |
+
|
| 44 |
+
# 4. Binarización Adaptativa
|
| 45 |
+
binary = cv2.adaptiveThreshold(
|
| 46 |
+
denoised, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2
|
| 47 |
+
)
|
| 48 |
+
return binary, img
|
| 49 |
+
except Exception as e:
|
| 50 |
+
print(f"Error en pre-procesamiento: {e}")
|
| 51 |
+
return None, img
|