| """ |
| Mejora F — Texto CLIP más informativo. |
| |
| El pipeline original usa los labels c-TF-IDF (4 keywords como |
| "lince_cría_ejemplares_hábitat") como texto para CLIP. Esto produce |
| una señal de solo 1.10× sobre baseline aleatorio. |
| |
| Este módulo proporciona funciones para seleccionar el headline más |
| representativo de cada micro-evento como texto CLIP, obteniendo |
| oraciones completas y naturales mejor alineadas con el espacio visual. |
| |
| Uso típico en Colab: |
| from src.multimodal.clip_utils import build_clip_text_inputs, compute_clip_coherence |
| text_inputs = build_clip_text_inputs(df_final, clip_model, clip_preprocess) |
| df_clip = compute_clip_coherence(df_final, text_inputs, image_embeddings) |
| """ |
|
|
| from __future__ import annotations |
|
|
| import logging |
| from typing import Literal, Optional |
|
|
| import numpy as np |
| import pandas as pd |
| from sklearn.metrics.pairwise import cosine_similarity |
|
|
| logger = logging.getLogger(__name__) |
|
|
| TextStrategy = Literal['best_headline', 'concat_headlines', 'label_jit'] |
|
|
|
|
| |
| |
| |
| def _centroid_representative_headline( |
| grupo: pd.DataFrame, |
| embeddings_matrix: np.ndarray, |
| ) -> str: |
| """ |
| Devuelve el headline del artículo más cercano al centroide semántico |
| del micro-evento (mejor representante del grupo). |
| """ |
| idx = grupo.index.tolist() |
| if not idx: |
| return '' |
| embs = embeddings_matrix[idx] |
| centroide = embs.mean(axis=0, keepdims=True) |
| sims = cosine_similarity(centroide, embs)[0] |
| best_local = int(np.argmax(sims)) |
| best_global = idx[best_local] |
| headline = grupo.loc[best_global, 'headline'] |
| return str(headline) if not pd.isna(headline) else '' |
|
|
|
|
| def build_event_texts( |
| df_final: pd.DataFrame, |
| embeddings_matrix: Optional[np.ndarray] = None, |
| strategy: TextStrategy = 'best_headline', |
| max_headlines: int = 3, |
| event_id_col: str = 'micro_event_id', |
| headline_col: str = 'headline', |
| ) -> dict[str, str]: |
| """ |
| Construye un diccionario {micro_event_id → texto} para usar como input de CLIP. |
| |
| Estrategias disponibles |
| ----------------------- |
| 'best_headline' : El headline del artículo más cercano al centroide del evento. |
| Requiere embeddings_matrix. |
| 'concat_headlines' : Concatenación de los max_headlines titulares más largos. |
| No requiere embeddings. |
| 'label_jit' : Usa el label c-TF-IDF actual (baseline, para comparación). |
| |
| Parámetros |
| ---------- |
| df_final : DataFrame con columnas micro_event_id y headline. |
| embeddings_matrix : Array (N_total, D) de embeddings (solo para 'best_headline'). |
| strategy : Estrategia de selección de texto. |
| max_headlines : Número máximo de headlines a concatenar ('concat_headlines'). |
| event_id_col : Nombre de la columna de ID de micro-evento. |
| headline_col : Nombre de la columna de titular. |
| |
| Retorna |
| ------- |
| dict {event_id: texto} |
| """ |
| texts: dict[str, str] = {} |
|
|
| for event_id, grupo in df_final.groupby(event_id_col): |
| if strategy == 'best_headline': |
| if embeddings_matrix is None: |
| raise ValueError("'best_headline' requiere embeddings_matrix") |
| texto = _centroid_representative_headline(grupo, embeddings_matrix) |
|
|
| elif strategy == 'concat_headlines': |
| headlines = ( |
| grupo[headline_col] |
| .dropna() |
| .sort_values(key=lambda s: s.str.len(), ascending=False) |
| .head(max_headlines) |
| .tolist() |
| ) |
| texto = '. '.join(h.strip().rstrip('.') for h in headlines if h.strip()) |
|
|
| elif strategy == 'label_jit': |
| label = grupo['label_jit'].dropna().iloc[0] if 'label_jit' in grupo.columns else '' |
| texto = str(label).replace('_', ' ') |
|
|
| else: |
| raise ValueError(f"Estrategia desconocida: {strategy!r}") |
|
|
| texts[str(event_id)] = texto |
|
|
| empty = sum(1 for t in texts.values() if not t.strip()) |
| logger.info( |
| f"Textos CLIP construidos ({strategy}): " |
| f"{len(texts)} eventos, {empty} vacíos ({empty/max(len(texts),1)*100:.1f}%)" |
| ) |
| return texts |
|
|
|
|
| |
| |
| |
| def encode_texts_clip( |
| event_texts: dict[str, str], |
| clip_model, |
| clip_tokenize, |
| device: str = 'cuda', |
| batch_size: int = 256, |
| ) -> dict[str, np.ndarray]: |
| """ |
| Codifica los textos de cada micro-evento con CLIP y devuelve |
| {event_id: embedding L2-normalizado (512,)}. |
| |
| Parámetros |
| ---------- |
| event_texts : Salida de build_event_texts(). |
| clip_model : Modelo CLIP (openai/clip). |
| clip_tokenize : clip.tokenize. |
| device : 'cuda' o 'cpu'. |
| batch_size : Artículos por batch (CLIP text encoder es rápido). |
| """ |
| import torch |
|
|
| event_ids = list(event_texts.keys()) |
| texts = [event_texts[eid] for eid in event_ids] |
| embeddings: dict[str, np.ndarray] = {} |
|
|
| clip_model.eval() |
|
|
| with torch.no_grad(): |
| for i in range(0, len(texts), batch_size): |
| batch_ids = event_ids[i:i + batch_size] |
| batch_txts = texts[i:i + batch_size] |
|
|
| |
| tokens = clip_tokenize(batch_txts, truncate=True).to(device) |
| feats = clip_model.encode_text(tokens) |
| feats = feats / feats.norm(dim=-1, keepdim=True) |
| feats = feats.cpu().numpy().astype(np.float32) |
|
|
| for eid, vec in zip(batch_ids, feats): |
| embeddings[eid] = vec |
|
|
| logger.info(f"Embeddings de texto CLIP generados: {len(embeddings)} eventos") |
| return embeddings |
|
|
|
|
| |
| |
| |
| def compute_clip_coherence( |
| df_resumen: pd.DataFrame, |
| text_embeddings: dict[str, np.ndarray], |
| image_embeddings_by_event: dict[str, np.ndarray], |
| event_id_col: str = 'micro_event_id', |
| ) -> pd.DataFrame: |
| """ |
| Calcula la coherencia coseno imagen-texto para cada micro-evento. |
| |
| Parámetros |
| ---------- |
| df_resumen : DataFrame resumen con una fila por micro-evento. |
| text_embeddings : {event_id: text_emb(512,)} de encode_texts_clip(). |
| image_embeddings_by_event: {event_id: image_emb_mean(512,)} de la Fase 3. |
| event_id_col : Nombre de la columna de ID. |
| |
| Retorna |
| ------- |
| df_resumen con columna 'clip_text_coherence_v2' añadida. |
| """ |
| coherences: list[float] = [] |
|
|
| for _, row in df_resumen.iterrows(): |
| eid = str(row[event_id_col]) |
| t_emb = text_embeddings.get(eid) |
| i_emb = image_embeddings_by_event.get(eid) |
|
|
| if t_emb is None or i_emb is None: |
| coherences.append(float('nan')) |
| continue |
|
|
| t = t_emb.reshape(1, -1) |
| im = i_emb.reshape(1, -1) |
| coh = float(cosine_similarity(t, im)[0, 0]) |
| coherences.append(coh) |
|
|
| df_out = df_resumen.copy() |
| df_out['clip_text_coherence'] = coherences |
|
|
| valid = [c for c in coherences if not np.isnan(c)] |
| if valid: |
| baseline = np.mean(valid) |
| real_m = np.mean(valid) |
| print(f"\nCoherencia CLIP (estrategia mejorada):") |
| print(f" N : {len(valid)}") |
| print(f" Media : {real_m:.4f}") |
| print(f" Mediana : {np.median(valid):.4f}") |
| print(f" Std : {np.std(valid):.4f}") |
|
|
| return df_out |
|
|
|
|
| |
| |
| |
| def compare_coherence_strategies( |
| df_resumen: pd.DataFrame, |
| text_emb_original: dict[str, np.ndarray], |
| text_emb_improved: dict[str, np.ndarray], |
| image_embeddings_by_event: dict[str, np.ndarray], |
| event_id_col: str = 'micro_event_id', |
| ) -> pd.DataFrame: |
| """ |
| Calcula la coherencia con ambas estrategias (label_jit vs best_headline) |
| y muestra una comparativa directa. |
| |
| Útil para validar cuánto mejora la señal imagen-texto antes de reemplazar |
| el pipeline completo. |
| """ |
| df1 = compute_clip_coherence(df_resumen, text_emb_original, image_embeddings_by_event, event_id_col) |
| df2 = compute_clip_coherence(df_resumen, text_emb_improved, image_embeddings_by_event, event_id_col) |
|
|
| c_orig = df1['clip_text_coherence'].dropna() |
| c_impr = df2['clip_text_coherence'].dropna() |
|
|
| |
| rng = np.random.default_rng(42) |
| event_ids = list(image_embeddings_by_event.keys()) |
| img_matrix = np.stack([image_embeddings_by_event[e] for e in event_ids]) |
| txt_improved = np.stack([text_emb_improved[e] for e in event_ids if e in text_emb_improved]) |
| if len(txt_improved) == len(img_matrix): |
| perm = rng.permutation(len(img_matrix)) |
| baseline = float(cosine_similarity(txt_improved, img_matrix[perm]).diagonal().mean()) |
| else: |
| baseline = float('nan') |
|
|
| print("\n" + "=" * 55) |
| print(" COMPARATIVA DE ESTRATEGIAS DE TEXTO CLIP") |
| print("=" * 55) |
| print(f" Baseline aleatorio (random) : {baseline:.4f}") |
| print(f" label c-TF-IDF (original) : {c_orig.mean():.4f} " |
| f"({c_orig.mean()/baseline:.2f}× baseline)") |
| print(f" Best headline (mejorado) : {c_impr.mean():.4f} " |
| f"({c_impr.mean()/baseline:.2f}× baseline)") |
| print("=" * 55) |
|
|
| return df2 |
|
|