Datasets:

pacoalberola
/

Poke-Grader-Card-Identifier

text stringlengths 7 13
en-base1-1
en-base1-2
en-base1-3
en-base1-4
en-base1-5
en-base1-6
en-base1-7
en-base1-8
en-base1-9
en-base1-10
en-base1-11
en-base1-12
en-base1-13
en-base1-14
en-base1-15
en-base1-16
en-base1-17
en-base1-18
en-base1-19
en-base1-20
en-base1-21
en-base1-22
en-base1-23
en-base1-24
en-base1-25
en-base1-26
en-base1-27
en-base1-28
en-base1-29
en-base1-30
en-base1-31
en-base1-32
en-base1-33
en-base1-34
en-base1-35
en-base1-36
en-base1-37
en-base1-38
en-base1-39
en-base1-40
en-base1-41
en-base1-42
en-base1-43
en-base1-44
en-base1-45
en-base1-46
en-base1-47
en-base1-48
en-base1-49
en-base1-50
en-base1-51
en-base1-52
en-base1-53
en-base1-54
en-base1-55
en-base1-56
en-base1-57
en-base1-58
en-base1-59
en-base1-60
en-base1-61
en-base1-62
en-base1-63
en-base1-64
en-base1-65
en-base1-66
en-base1-67
en-base1-68
en-base1-69
en-base1-70
en-base1-71
en-base1-72
en-base1-73
en-base1-74
en-base1-75
en-base1-76
en-base1-77
en-base1-78
en-base1-79
en-base1-80
en-base1-81
en-base1-82
en-base1-83
en-base1-84
en-base1-85
en-base1-86
en-base1-87
en-base1-88
en-base1-89
en-base1-90
en-base1-91
en-base1-92
en-base1-93
en-base1-94
en-base1-95
en-base1-96
en-base1-97
en-base1-98
en-base1-99
en-base1-100

End of preview. Expand in Data Studio

🎴 PokeGrader: Card Identifier Index (DINOv2 + FAISS)

Este repositorio contiene el índice de búsqueda vectorial para identificar cartas Pokémon. Utiliza DINOv2 (facebook/dinov2-base) para la extracción de características y FAISS (IndexFlatIP) para búsqueda por cosine similarity en ~15ms.

📊 Métricas

Métrica	Valor
Recall@1	100.0%
Cartas indexadas	5000

Recall@1 calculado sobre 500 consultas aleatorias del propio índice.

🗂️ Arquitectura

Componente	Detalle
Backbone	DINOv2 (`facebook/dinov2-base`, ViT-B/14)
Dimensión embedding	768
Índice FAISS	`IndexFlatIP` (cosine sim tras L2-norm)
Latencia búsqueda	~15ms para 5000 cartas

📂 Contenido

Archivo	Descripción
`card_index.faiss`	Índice binario de búsqueda vectorial
`card_id_map.json`	Mapeo posición → card_id
`catalog_indexed.csv`	Metadatos de cada carta (nombre, set, etc.)
`embeddings.npy`	Vectores crudos para regenerar el índice

📝 Notas

Entrenado con 500 cartas, Resultado: Índice FAISS construido: 5000 vectores, dim=768 Top-5 similares a la primera carta: 1. en-base1-1 (sim=1.0000) 2. en-lc-1 (sim=0.9411) 3. en-base4-1 (sim=0.9155) 4. en-base5-1 (sim=0.8131) 5. en-base3-6 (sim=0.7952)

🚀 Uso en producción

import faiss, json, numpy as np
from transformers import AutoImageProcessor, AutoModel
from PIL import Image
import torch

processor  = AutoImageProcessor.from_pretrained('facebook/dinov2-base')
dino       = AutoModel.from_pretrained('facebook/dinov2-base').eval()
index      = faiss.read_index('card_index.faiss')
card_ids   = json.load(open('card_id_map.json'))

img    = Image.open('query.jpg').convert('RGB')
inputs = processor(images=[img], return_tensors='pt')
with torch.no_grad():
    emb = dino(**inputs).last_hidden_state[:, 0, :].numpy().astype(np.float32)
faiss.normalize_L2(emb)
D, I = index.search(emb, k=5)
print([card_ids[i] for i in I[0]])

Downloads last month: 105