MNIST MLP Classifier
This is a fully connected MNIST digit classifier trained as part of the
dlab deep-learning experimentation roadmap.
The model is intentionally an MLP rather than a CNN. It is useful as a strong baseline for studying optimization, regularization, augmentation, seed variance, and the ceiling of non-convolutional models on MNIST.
Results
10-seed confirmation sweep:
| metric | result |
|---|---|
| validation accuracy | 99.3600% ± 0.0817 pp |
| validation loss | 0.15172 ± 0.00235 |
| test accuracy | 99.4470% ± 0.0195 pp |
| test loss | 0.14746 ± 0.00034 |
| test errors | 55.3 ± 1.95 / 10000 |
Architecture
- Dataset: MNIST
- Model: MLP
- Hidden width: 1024
- Hidden layers: 3
- Activation: ReLU
- Batch normalization: enabled
- Dropout: 0.2
- Optimizer: Adam
- Learning rate: 0.001
- Scheduler: OneCycleLR
- Weight decay: 0.0001
- Label smoothing: 0.02
- Weight averaging: EMA
- Train augmentation: random affine, 10 degrees, translate 0.05, scale 0.9-1.1
Files
model.ckpt: PyTorch Lightning checkpoint from the best run.model.onnx: ONNX export of the EMA/current model state used for evaluation.config.yaml: resolved Hydra training config.metrics.csv: training metrics from the run.metadata.json: compact metadata for inference and provenance.
Preprocessing
Inputs should be MNIST grayscale images converted to tensors and normalized with:
mean = [0.1307]
std = [0.3081]
The ONNX model expects float tensors shaped [batch, 1, 28, 28] under the input
name images, and returns class logits under the output name logits.
Provenance
- Downloads last month
- 23