convnext_tiny-stanford-cars-brand

Model klasifikasi gambar mobil hasil fine-tuning convnext_tiny (pretrained ImageNet) pada dataset tanganke/stanford_cars.

• Arsitektur : convnext_tiny (CNN)
• Jumlah kelas : 49 (mode label: brand)
• Ukuran input : 288x288
• Epoch : 8
• Akurasi test : 92.70% (baseline kelas mayoritas: 11.11%)

Cara pakai

import torch, torch.nn as nn
from huggingface_hub import hf_hub_download
from torchvision import transforms
from torchvision.models import resnet50, convnext_tiny
from PIL import Image

path = hf_hub_download("ryosiswand/convnext_tiny-stanford-cars-brand", "model.pth")
ckpt = torch.load(path, map_location="cpu", weights_only=False)

# bangun arsitektur sesuai ckpt["arch"], lalu muat bobotnya
if ckpt["arch"] == "convnext_tiny":
    model = convnext_tiny(weights=None)
    model.classifier[2] = nn.Linear(model.classifier[2].in_features, ckpt["num_classes"])
else:
    model = resnet50(weights=None)
    model.fc = nn.Linear(model.fc.in_features, ckpt["num_classes"])
model.load_state_dict(ckpt["state_dict"])
model.eval()

tf = transforms.Compose([
    transforms.Resize(int(ckpt["img_size"]*1.14)), transforms.CenterCrop(ckpt["img_size"]), transforms.ToTensor(),
    transforms.Normalize(ckpt["mean"], ckpt["std"]),
])
img = Image.open("mobil.jpg").convert("RGB")
probs = model(tf(img).unsqueeze(0)).softmax(-1)[0]
print(ckpt["class_names"][int(probs.argmax())], float(probs.max()))

Dibuat sebagai proyek Computer Vision (CNN) - Ruangguru AI Bootcamp.