SADIK Vision Multiverse - TinyNet (54 KB)

A very small int8 image classifier for the 4-class SADIK Vision Multiverse dataset, created by Sadik Nawaz (Age 11).

• Weights file: tinynet_int8.bin - 53.60 KB (54,882 bytes), single self-contained blob
• Parameters: 47,236
• Input: 96x96 RGB
• Backbone: depthwise-separable CNN trained from scratch via knowledge distillation from the 83%-accuracy ResNet18 teacher
• Quantization: int8 per-output-channel symmetric weights, float16 BatchNorm/bias, float16 scales

Goal: a model that fits in under 59 KB while still learning the 4 classes.

---

Results (int8, 5-fold cross-validation)

The dataset has only 77 hand-cleaned images, so treat these numbers as an educational estimate, not a production benchmark. The tiny 54 KB model is intentionally much weaker than the 83% ResNet18 teacher.

---

Usage

pip install torch torchvision numpy pillow
python predict_tiny.py my_image.jpg

from tinynet import load_int8, read_extra
from PIL import Image
from torchvision import transforms
import torch

model = load_int8("tinynet_int8.bin")
info = read_extra("tinynet_int8.bin")
tf = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.Resize((96, 96)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
with torch.no_grad():
    probs = model(tf(Image.open("my_image.jpg").convert("RGB")).unsqueeze(0)).softmax(1)[0]
for cls, p in sorted(zip(info["class_names"], probs.tolist()), key=lambda x: -x[1]):
    print(f"{cls:28s} {p:.1%}")

The int8 file is dequantized to float32 at load time, so it runs anywhere PyTorch runs - no special int8 kernels needed.

---

Training

1. Train a ResNet18 teacher (5-fold CV, 83.1% out-of-fold).
2. Cache 16 augmented views per image plus the teacher's soft logits.
3. Distill the teacher into TinyNet (temperature 2.0, 50/50 cross-entropy + KL loss).
4. Quantize weights to int8 per output channel.
5. Evaluate on held-out folds (61.0% int8 out-of-fold accuracy).

Files:

• tinynet_int8.bin - 53.6 KB quantized weights (the tiny model)
• tinynet_fp32.pt - 214.1 KB float32 checkpoint
• tinynet.py - architecture + quantization/loading code
• fast_train.py - full training + distillation script
• predict_tiny.py - CLI inference
• metrics.json, sizes.json

---

Limitations

• 77 training images only; accuracy is noisy.
• 96x96 input, so fine details are lost.
• cyberpunk_tech vs game_assets_3d is the hardest pair.
• Int8 weights are dequantized for inference (small file, not faster compute).
• Not suitable for safety-critical or production use.

---

Author

Sadik Nawaz - 11-year-old AI developer from Pakistan

• Hugging Face: @SadikNawazAI