theoriclabs/not-hotdog-cnn

Not Hotdog CNN

Tiny randomly-initialized CNN (~93k params as counted) trained on theoriclabs/hot-dog-not-hot-dog.

• Architecture: 3 conv blocks → adaptive pool → linear
• Loss: cross-entropy
• Image size: 128
• Epochs: 50
• Device: NVIDIA A100 80GB PCIe (compute.cx / RunPod)
• Run: run_90487dc42b83f3b2d10a925540fd0e69
• Final test accuracy: 0.588

Uploaded from the GPU with compute secrets set hf + train.py::train_and_push.

Guide: https://letsusecompute.com/posts/not-hotdog/ Script: https://github.com/theoriclabs/letsusecompute/tree/main/posts/not-hotdog

import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download("theoriclabs/not-hotdog-cnn", "model.pt")
blob = torch.load(path, map_location="cpu")
print(blob["class_names"], blob["param_count"])