CIFAR-10 CNN — 🤖 Fully AI-Made
This model — the architecture, training code, evaluation, plots, and this very card — was created end-to-end by an AI agent (Claude, via the Hermes ML research loop). No human wrote the code. Kaggle T4 GPU did the compute; the artifact was pushed to the Hugging Face Hub automatically.
A compact 3-block VGG-style CNN trained from scratch on CIFAR-10.
Results
| metric | value |
|---|---|
| Test accuracy | 91.52% |
| Test loss | 0.542 |
| Best val accuracy | 91.52% |
| Parameters | 1.15M |
| Epochs | 18 |
| Train time (T4) | 5.9 min |
Training curves
Confusion matrix
Per-class accuracy
| class | accuracy |
|---|---|
| automobile | 96.0% |
| truck | 95.2% |
| frog | 94.9% |
| ship | 94.6% |
| horse | 93.7% |
| airplane | 92.7% |
| deer | 91.7% |
| dog | 86.5% |
| bird | 86.5% |
| cat | 83.4% |
Sample predictions
Architecture
3 conv blocks (64→128→256 ch, each = 2×[Conv3×3–BN–ReLU]+MaxPool) → GAP → Dropout(0.3) → Linear(10). AdamW + OneCycleLR, AMP, label smoothing 0.05, random crop + h-flip augmentation.
Usage
import torch, torch.nn as nn, torch.nn.functional as F
# (define the CNN class as in config.json), then:
sd = torch.load("cifar10_cnn.pt", map_location="cpu")
model.load_state_dict(sd); model.eval()
Provenance
- Dataset:
uoft-cs/cifar10 - Compute: Kaggle Notebook, 1× NVIDIA Tesla T4
- Pipeline: Hermes
ml-research-loop(Kaggle GPU → Hugging Face) - Author: AI agent (Claude). Human involvement: pressed go.