SadikNawazAI/sadik-vision-multiverse-classifier

🌌 SADIK Vision Multiverse Classifier

A ResNet18 image classifier fine-tuned on the SADIK Vision Multiverse Dataset, created by Sadik Nawaz (Age 11).

The model classifies images into 4 categories:

ID	Class	Meaning
0	anime_digital_art	Anime, manga & digital art
1	cyberpunk_tech	Cyberpunk, neon cities, robots, cyborgs
2	game_assets_3d	3D renders, low-poly game assets
3	photorealistic_portraits	Real photo portraits

Note: class order is alphabetical, matching torchvision.datasets.ImageFolder.

---

📊 Results (honest 5-fold cross-validation)

Metric	Score
Out-of-fold accuracy	83.1%
5-fold CV accuracy	83.3% ± 6.3%
Out-of-fold macro F1	77.8%

Class	Precision	Recall	F1	Support
anime_digital_art	0.73	0.73	0.73	11
cyberpunk_tech	0.71	0.83	0.77	24
game_assets_3d	0.75	0.55	0.63	11
photorealistic_portraits	1.00	0.97	0.98	31

The dataset has only 77 hand-cleaned images, so these numbers are noisy. The confusion matrix and training curves are in this repo.

---

🚀 Usage

import torch
from PIL import Image
from torchvision import transforms
from torchvision.models import resnet18

ckpt = torch.load("model.pt", map_location="cpu")
model = resnet18(weights=None)
model.fc = torch.nn.Linear(model.fc.in_features, len(ckpt["class_names"]))
model.load_state_dict(ckpt["state_dict"])
model.eval()

tf = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

image = Image.open("your_image.jpg").convert("RGB")
with torch.no_grad():
    probs = model(tf(image).unsqueeze(0)).softmax(1)[0]

for cls, p in sorted(zip(ckpt["class_names"], probs.tolist()), key=lambda x: -x[1]):
    print(f"{cls:28s} {p:.1%}")

Or use the included CLI:

python predict.py path/to/image1.jpg path/to/image2.jpg

---

🏋️ Training Details

• Backbone: ResNet18 (ImageNet pretrained)
• Input: 224×224, ImageNet normalization
• Augmentation: RandomResizedCrop, horizontal flip, rotation, color jitter, RandAugment, random erasing
• Loss: CrossEntropy with class weights + label smoothing 0.1
• Optimizer: AdamW (lr 3e-4, weight decay 1e-4), cosine schedule
• Validation: stratified 5-fold cross-validation
• Final model: trained on all 77 images for 8 epochs
• Hardware: CPU only

Reproduce with:

python train.py --model resnet18 --data-dir ./data --out-dir ./artifacts

---

⚠️ Limitations

• Trained on a very small educational dataset (77 images).
• Not suitable for safety-critical or production use.
• Performance may be lower on images that look different from the training set.
• Some classes are visually similar (cyberpunk_tech vs game_assets_3d), which causes most mistakes.

---

👨‍💻 Author

Sadik Nawaz — 11-year-old AI developer from Pakistan 🇵🇰

• Hugging Face: @SadikNawazAI
• Dataset: SADIK Vision Multiverse