Wisnu1354/Gumi-Banten

🌿 Gumi Banten Plant Identification — CNN + ViT Hybrid

Model identifikasi tanaman Gumi Banten menggunakan arsitektur hybrid CNN (EfficientNet-B4) + Vision Transformer (ViT).

📊 Performa Model

Metrik	Nilai
Test Accuracy	0.9012 (90.12%)
Best Val Accuracy	0.9202 (92.02%)
Jumlah Kelas	10

🏗️ Arsitektur

• CNN Backbone: EfficientNet-B4 (feature extractor)
• ViT Head: 6x Transformer Encoder Block, embed_dim=384
• Input Size: 224×224 px
• Framework: PyTorch (weights disimpan sebagai HDF5/.h5)

📦 File dalam Repo

File	Deskripsi
gumi_banten_cnn_vit.h5	Model weights dalam format HDF5
best_model.pth	Checkpoint PyTorch lengkap
config.json	Konfigurasi model
class_names.txt	Daftar nama kelas

🚀 Cara Menggunakan

Load dari .pth (PyTorch — Direkomendasikan)

import torch
from huggingface_hub import hf_hub_download

pth_path = hf_hub_download(repo_id="Wisnu1354/Gumi-Banten", filename="best_model.pth")
ckpt = torch.load(pth_path, map_location='cpu')
model.load_state_dict(ckpt['model_state'])
model.eval()

Load dari .h5

import h5py, torch, numpy as np
from huggingface_hub import hf_hub_download

def load_from_h5(h5_path, model_class, cfg):
    with h5py.File(h5_path, 'r') as hf:
        class_names = list(hf['metadata/class_names'][:])
        state_dict = {}
        def _load(name, obj):
            if isinstance(obj, h5py.Dataset):
                state_dict[name.replace('/', '.')] = torch.tensor(obj[()])
        hf['model_weights'].visititems(_load)
    model = model_class(cfg)
    model.load_state_dict(state_dict)
    model.eval()
    return model, class_names

h5_path = hf_hub_download(repo_id="Wisnu1354/Gumi-Banten", filename="gumi_banten_cnn_vit.h5")
model, class_names = load_from_h5(h5_path, CNNViTHybrid, CFG)

📋 Kelas yang Didukung

• Daun Ancak
• Daun Base
• Daun Bila
• Daun Bingin
• Daun Dapdap
• Daun Intaran
• Daun Kayu Tulak
• Daun Kelor
• Daun Nagasari
• Daun Pucuk Rejuna

🧑‍💻 Training

Dilatih di Google Colab menggunakan GPU A100/V100/T4.

---

Dibuat oleh: Gumi Banten Research | 2026-04-27